AgentScout Logo Agent Scout

AI Agent Weekly Intel (W33): MCP Tunnels, Valuation Club, Memory Architecture Shift

MCP tunnels enable private enterprise access without firewall changes through outbound-only gateways. The $10B+ valuation club reveals distinct models: Cursor 25x ARR, Cognition 73x growth, Sierra enterprise CX. Observational memory cuts costs 10x vs RAG.

AgentScout · ·
#mcp-tunnels #ai-agents #memory-architecture #enterprise-deployment #valuation #protocols
Analyzing Data Nodes...
SIG_CONF:CALCULATING
Verified Sources

TL;DR

Three converging threads define this week’s AI agent ecosystem: MCP tunnels fundamentally restructure enterprise deployment economics by eliminating inbound firewall requirements through outbound-only gateway architecture; a tiered $10B+ valuation club emerges with Cursor ($50B, 25x ARR), Cognition ($25B, 73x growth), Sierra ($15.8B), and Lovable ($6.6B) showing distinct business model patterns; and observational memory architecture displaces RAG on long-context benchmarks (94.87% vs 80.05%) while cutting token costs 10x.

Key Facts

  • Who: Anthropic (MCP tunnels), Cognition ($25B valuation target), Mastra (observational memory), Linux Foundation (A2A governance)
  • What: Infrastructure security maturation, valuation acceleration, memory architecture evolution, protocol competition
  • When: May 19-24, 2026 (MCP tunnels launch), April-June 2026 (valuation announcements), Q1 2026 (hybrid retrieval tripling)
  • Impact: 150+ organizations on A2A protocol, 12% enterprise production rate, 33.3% hybrid retrieval adoption intent

Executive Summary

The AI agent ecosystem underwent three structural shifts this week that will shape enterprise deployment decisions for the next 18 months. First, Anthropic’s launch of MCP tunnels at ‘Code with Claude’ London on May 19, 2026, introduces an outbound-only gateway architecture that eliminates the need for inbound firewall rules—agents access private databases and internal APIs through a lightweight gateway running inside the enterprise network that establishes encrypted mTLS connections to Anthropic infrastructure. This architecture removes the single largest barrier to enterprise adoption: security team approvals for perimeter changes.

Second, the $10B+ valuation club crystallized into four distinct business model patterns: Cursor’s developer IDE model ($50B valuation, 25x ARR multiple, 67% Fortune 500 penetration), Cognition’s autonomous engineer approach ($25B target, 73x ARR growth from $1M to $73M in nine months), Sierra’s enterprise customer experience platform ($15.8B post-money after $950M raise), and Lovable’s vibe coding for non-technical users ($6.6B valuation, 15 million daily active users). Each targets a different segment with fundamentally different unit economics.

Third, observational memory architecture emerged as a viable alternative to RAG for long-running agent workloads. Mastra’s open-source implementation achieved 94.87% on the LongMemEval benchmark with GPT-5-mini (84.23% with GPT-4o), surpassing RAG baseline (80.05%) while reducing token costs by a factor of 10. Redis entered this space with the Iris platform launch on May 18, 2026, offering Context Retriever and Agent Memory components that run 99% of data on SSD at one-tenth the cost of in-memory storage. Enterprise adoption intent for hybrid retrieval tripled from 10.3% to 33.3% in Q1 2026, signaling that standalone RAG programs are hitting scale walls.

The protocol layer also saw consolidation. A2A (Agent-to-Agent) reached 150+ organizations in production at its one-year mark, while MCP was donated to the Linux Foundation’s Agentic AI Foundation (AAIF) with Anthropic, Block, and OpenAI as co-founders. These protocols are complementary: MCP standardizes tool and data connectivity (showing ROI in 2-3 months), while A2A enables agent-to-agent coordination (taking 6-12 months for full value). Enterprises should plan for dual-protocol deployments.

Background & Context

The AI agent infrastructure landscape evolved rapidly from experimental prototypes in 2024 to production deployments in 2026. Three forces drove this acceleration: the emergence of standardized protocols (MCP in December 2025, A2A in May 2026), the validation of autonomous coding agents as enterprise tools (Cognition’s Devin reaching $73M ARR), and the maturation of security frameworks (Five Eyes alliance publishing guidance on agent governance in April 2026).

Enterprise adoption patterns shifted from isolated pilot projects to integrated workflow deployments. The PwC-Anthropic alliance expansion on May 14, 2026—training and certifying 30,000 professionals on Claude—signals that professional services firms now view agentic technology not as experimental but as a core capability requiring systematic workforce development. G2 research indicates enterprises expect 10-25% of workflows to be managed by agents within months.

The memory architecture problem remained unsolved. RAG (Retrieval-Augmented Generation) programs hit scalability limits as context windows expanded but retrieval accuracy degraded. Hybrid retrieval adoption intent tripled in Q1 2026, reflecting a search for alternatives. Observational memory—using background agents to compress and maintain stable context windows—emerged as a viable solution for long-running agents, though not for dynamic knowledge base scenarios.

Infrastructure Security Maturation: MCP Tunnels

Outbound-Only Gateway Architecture

MCP tunnels represent a fundamental shift in how AI agents access private enterprise resources. The architecture operates on a simple principle: instead of requiring enterprises to open inbound firewall ports for agent access, a lightweight gateway runs inside the enterprise network and establishes an outbound encrypted mTLS connection to Anthropic’s tunnel control plane.

The security implications are substantial. Traditional agent deployments required enterprises to expose internal APIs and databases to the public internet, creating attack surfaces that security teams resisted. MCP tunnels eliminate this exposure by keeping all private resources behind the enterprise perimeter—the agent never holds credentials, and the gateway authenticates to the tunnel control plane.

Deployment Pattern:

ComponentLocationFunction
MCP GatewayInside enterprise networkOpens outbound mTLS connection to Anthropic
Tunnel Control PlaneAnthropic infrastructureRoutes agent requests through gateway
Private MCP ServersBehind enterprise firewallInternal databases, APIs accessed via tunnel
Agent SessionAnthropic infrastructureNo direct access to private resources

Source: Claude API Docs - MCP Tunnels Security

Enterprise Deployment Economics

The economic impact of removing firewall change requests from the deployment timeline is difficult to overstate. In most enterprises, firewall rule changes require approval from security, networking, and compliance teams—a process measured in weeks, not days. MCP tunnels reduce this timeline from “wait for approval” to “deploy gateway and configure outbound proxy.”

Thomson Reuters’ Legal MCP integration provides a concrete example. The deployment provides access to 1.9 billion Westlaw and Practical Law documents, 1.4 billion KeyCite validity signals, and a patent-pending citation ledger—all through MCP without requiring perimeter changes. For enterprises with similar internal knowledge bases, MCP tunnels represent immediate deployment enablement.

Governance Framework Requirements:

The Five Eyes alliance guidance published April 30, 2026, identifies autonomous agent behavior as a new enterprise attack surface requiring visibility and real-time controls. MCP tunnels address credential control at the network boundary, but enterprises must still implement:

  • Least-privilege role assignments for agent actions
  • RBAC (Role-Based Access Control) integration
  • Context limits to prevent runaway token consumption
  • Policy enforcement points at the gateway level
  • Continuous monitoring of agent-initiated actions

Source: Five Eyes ‘Careful Adoption of Agentic AI Services’

What This Means for Enterprises

For Security Teams: MCP tunnels reduce perimeter risk but shift security responsibility to identity governance and monitoring. The outbound-only model aligns with zero-trust principles—verify every request, never trust the network.

For Infrastructure Teams: Deployment velocity increases by an order of magnitude. A gateway can be deployed in hours, not weeks. The trade-off is dependency on Anthropic’s tunnel control plane—enterprises should negotiate SLAs and failover procedures.

For Procurement: The infrastructure conversation shifts from “we need firewall changes” to “we need gateway capacity planning.” Budget allocation moves from network perimeter expansion to internal gateway scaling.

Valuation Acceleration: The $10B+ Club

The AI agent market now has four distinct valuation tiers with fundamentally different business models. Understanding these models is essential for enterprise buyers evaluating build vs. buy decisions and for investors assessing market positioning.

Cursor: Developer IDE ($50B Valuation)

Cursor’s $50B valuation on $2B ARR (as of February 2026) represents a 25x multiple—exceptionally high for a developer tools company. The valuation reflects 67% Fortune 500 penetration and projected $6B ARR by year-end. Cursor 3, launched April 2026, introduced parallel agents and multi-session interfaces, directly competing with Claude Code’s desktop redesign.

Business Model: Enterprise subscriptions and developer seats. Revenue driver shifted from individual developers to enterprise contracts.

Key Metric: 67% Fortune 500 penetration indicates enterprise adoption validation.

Source: TechCrunch - Cursor $50B Valuation

Cognition: Autonomous Software Engineer ($25B Valuation Target)

Cognition’s ARR trajectory is unprecedented: $1M (September 2024) to $73M (June 2025)—73x growth in nine months. The $25B valuation target, if achieved, would more than double the $10.2B valuation from September 2025. Devin, the autonomous software engineer, can be pointed at a Jira ticket and will write code, run tests, and open PRs with minimal human intervention.

Business Model: Subscription to autonomous coding agent. Customers pay for engineering capacity without hiring.

Key Metrics: $1M to $73M ARR in 9 months, total net burn under $20M across company history (capital-efficient growth), Windsurf acquisition ($250M) for IDE capabilities.

Enterprise Partnerships: Infosys (first large digital services firm to deploy Devin at scale), Cognizant (AI-driven development collaboration).

Source: Cognition Official Blog

Sierra: Enterprise Customer Experience ($15.8B Valuation)

Sierra raised $950M at $15.8B post-money valuation—a 60% increase from the previous $10B round. Founded by Bret Taylor (OpenAI Chairman, former Salesforce Co-CEO), Sierra positions as “global standard” for AI-powered customer experiences. Unlike narrow support chatbots, Sierra agents handle full customer lifecycle: support, sales, account management.

Business Model: “Productized BPO” (Business Process Outsourcing). Sierra doesn’t just sell software—it takes on customer service operations.

Key Metric: $950M raise led by Tiger Global and GV, total capital now exceeds $1B.

Source: Sierra Official Blog

Lovable: Vibe Coding for Non-Technical Users ($6.6B Valuation)

Lovable targets the largest addressable market: non-technical users who want to build software without learning to code. The platform has 15 million daily active users and 200,000 new projects created daily. Projected $200M ARR for 2026.

Business Model: Subscriptions for non-technical creators. Lower per-seat price but massive volume.

Key Metrics: 15M DAU, 200K daily projects, $200M+ projected ARR (33x multiple at $6.6B).

Valuation Club Comparison

CompanyValuationARRARR MultipleTarget SegmentKey Differentiator
Cursor$50B$2B (Feb 2026)25xDevelopersIDE integration, Fortune 500 penetration
Cognition$25B$73M (Jun 2025)342x (implied)Engineering teamsAutonomous agent, 73x growth
Sierra$15.8BNot disclosedN/AEnterprise CXFull lifecycle, BPO model
Lovable$6.6B$200M+ (2026 proj)33xNon-technical usersVibe coding, 15M DAU

Strategic Insight: The highest multiples (Cursor 25x, Lovable 33x) attach to platforms with demonstrated enterprise penetration or massive user bases. Cognition’s implied 342x multiple on current ARR reflects growth expectations, not current revenue. Sierra’s valuation reflects founder pedigree and enterprise CX market size, not ARR multiple.

Memory Architecture Evolution: Observational Memory vs. RAG

The RAG paradigm—retrieve relevant documents, inject into context window, generate response—is hitting fundamental limits as agent workloads become more sophisticated. Three problems dominate: retrieval accuracy degrades as knowledge bases expand, context windows consume tokens at unsustainable rates, and long-running agents accumulate context that overwhelms retrieval systems.

Observational memory addresses these problems through a different architecture: background agents continuously compress and curate context, maintaining stable windows (~30k tokens) that provide relevant information without retrieval overhead.

Benchmark Results: LongMemEval

Mastra’s open-source observational memory implementation achieved 94.87% on LongMemEval with GPT-5-mini, surpassing RAG baseline (80.05%) by 14.82 percentage points. With GPT-4o, observational memory scored 84.23%—still 4.18 points above RAG.

Architecture Components:

ComponentFunctionToken Impact
Observer AgentMonitors all interactions, extracts salient informationAdds minimal overhead
Reflector AgentGarbage collects irrelevant observations when threshold (~40k tokens) reachedMaintains ~30k stable window
Three-Tier RepresentationProgressively compresses informationPrioritizes critical context

Cost Impact: 10x reduction in token costs. Stable context windows enable aggressive caching, while RAG requires fresh retrieval queries for each interaction.

Source: Mastra Research - Observational Memory

When to Choose: Observational Memory vs. RAG vs. Hybrid

ArchitectureBest ForCost ProfileAccuracy (LongMemEval)
Observational MemoryLong-running agents with stable context needs10x reduction94.87% (GPT-5-mini)
RAG BaselineDynamic knowledge bases, changing informationStandard retrieval costs80.05%
Hybrid RetrievalEnterprise RAG hitting scale limitsContext-dependentNot benchmarked

Enterprise Adoption Signal: Hybrid retrieval adoption intent tripled from 10.3% to 33.3% in Q1 2026 (VentureBeat VB Pulse data), indicating enterprises are moving away from standalone RAG toward context architectures.

Source: VentureBeat - Context Architecture Replacing RAG

Redis Iris Platform Entry

Redis launched the Iris platform on May 18, 2026, with five components targeting agent memory and context management:

  1. Context Retriever: Makes external data navigable by agents
  2. Agent Memory: Persistent context storage
  3. Redis Data Integration: Connects enterprise data sources
  4. LangCache: Caching layer for LLM responses
  5. Redis Search: Semantic search over context

Cost Efficiency: Redis Flex runs 99% of data on SSD at one-tenth the cost of in-memory storage, making large-scale context storage economically viable.

Source: Redis Official Blog

Protocol Competition: A2A vs. MCP

The agent interoperability layer is crystallizing around two protocols that serve different purposes. Understanding when to adopt each is critical for enterprise architecture decisions.

Protocol Comparison

DimensionMCPA2A
PurposeTool/data connectivityAgent-to-agent coordination
GovernanceLinux Foundation AAIFLinux Foundation AAIF
AdoptionUniversal (Anthropic standard)150+ organizations in production
ROI Timeline2-3 months6-12 months
Enterprise SupportAnthropic, Block, OpenAI, Google, Microsoft, AWSGoogle, Microsoft, AWS, Salesforce, SAP, ServiceNow, IBM
Use Case”How do agents interact with tools?""How do agents work together?”

Source: TrueFoundry - MCP vs A2A, Linux Foundation - A2A Milestone

Governance Consolidation

Both protocols now fall under the Linux Foundation’s Agentic AI Foundation (AAIF), donated by Anthropic in December 2025 with Anthropic, Block, and OpenAI as co-founders. Google, Microsoft, AWS, Cloudflare, and Bloomberg support the foundation.

This consolidation signals that the protocol wars are ending—enterprises can safely adopt both without fear of fragmentation. The protocols are complementary:

  • MCP: Adopt for every data source and tool you want agents to access. Shows immediate ROI through automation acceleration.
  • A2A: Adopt when operating multiple AI systems that need to coordinate workflows. Requires more architectural planning.

Security Note: Neither protocol was designed with adversarial input as a first-class concern. Enterprises must layer additional security controls—input validation, output filtering, behavior monitoring—on top of protocol-level trust.

Enterprise Deployment Patterns

Production Metrics: What Distinguishes Successful Deployments

Enterprise data from Digital Applied shows 12% of AI agent projects reach production. Successful deployments share four attributes:

  1. Pre-deployment infrastructure investment: Security, identity, monitoring built before agent deployment
  2. Governance documentation: Clear policies on agent authority, escalation procedures, audit trails
  3. Baseline metrics: Measurement frameworks established before deployment to demonstrate ROI
  4. Dedicated business ownership: Agents have accountable owners, not just technical custodians

Source: Digital Applied - Enterprise Adoption Data

Claude Code Parallel Agent Workflows

The Claude Code desktop redesign (April 14, 2026) introduced production-grade parallel agent management:

Key Features:

  • Multi-session sidebar: Manage active and recent sessions
  • Git worktree isolation: Each session scoped to .claude/worktrees/ directory, zero context bleed
  • Three view modes: Verbose, Normal, Summary
  • Cloud-based Routines: Scheduled automations running on Anthropic infrastructure

Agent View: Command center for steering multiple agents across parallel tasks. Emerging data shows 3-5 parallel sessions typical for complex projects.

Source: Claude Official Blog - Desktop Redesign

PwC-Anthropic Alliance: Enterprise Transformation Blueprint

The expanded alliance (May 14, 2026) establishes patterns for enterprise-wide agent deployment:

Three Focus Areas:

  1. Agentic technology build: Helping engineering teams build AI agent tools for clients
  2. AI-native deal-making: Deploying AI across M&A and deal processes
  3. Enterprise function reinvention: Reinventing operating models with AI

Scale: Joint Center of Excellence, 30,000 PwC professionals to be trained and certified on Claude. Claude-native finance business group launched.

ROI Patterns: Enterprises report 3x-5x ROI within 12-18 months for aggressive deployments. The critical success factor is workflow reimagination—redesigning processes around agent capabilities—not simply adding agents to existing workflows.

Source: PwC Official - Anthropic Alliance Expansion

Key Data Points

MetricValueSourceDate
Cognition ARR Growth73x ($1M to $73M in 9 months)Cognition Official BlogJun 2025
Observational Memory LongMemEval Score94.87% (GPT-5-mini)Mastra ResearchMay 2026
Observational Memory vs RAG+14.82 pts (94.87% vs 80.05%)Mastra ResearchMay 2026
Token Cost Reduction10xVentureBeatMay 2026
Hybrid Retrieval Adoption Intent10.3% → 33.3% (Q1 2026)VentureBeat VB PulseMar 2026
A2A Protocol Adoption150+ organizations in productionLinux FoundationMay 2026
Cursor Valuation$50B (25x ARR)TechCrunchApr 2026
Cursor Fortune 500 Penetration67%TechCrunchApr 2026
Sierra Valuation$15.8B post-moneyTechCrunchMay 2026
Lovable Daily Active Users15 millionPanto AI AnalysisApr 2026
Enterprise Production Rate12%Digital AppliedQ1 2026
PwC Claude Certification Target30,000 professionalsPwC OfficialMay 2026
Redis Flex Cost Efficiency99% data on SSD at 1/10th costRedis OfficialMay 2026
MCP Tunnels Launch’Code with Claude’ LondonAnthropic OfficialMay 19, 2026

Timeline: AI Agent Ecosystem (December 2025 - May 2026)

DateEventSignificance
Dec 2025Anthropic donates MCP to Linux Foundation AAIFInfrastructure standardization milestone
Dec 2025Cognition acquires Windsurf ($250M)AI IDE market consolidation
Apr 2026Cursor 3 launch with parallel agentsCompetitive response to Claude Code
Apr 14, 2026Claude Code desktop redesignProduction-grade parallel workflows
Apr 17, 2026Cursor $50B valuation talks$10B+ club tier formation
Apr 23, 2026Cognition $25B valuation talksAutonomous engineer market validation
Apr 30, 2026Five Eyes publishes agent governance guidanceGovernment-level security framework
May 1, 2026A2A protocol: 150+ organizations in productionProtocol adoption milestone
May 4, 2026Sierra raises $950M at $15.8BEnterprise CX market maturation
May 14, 2026PwC-Anthropic alliance expansionEnterprise deployment acceleration
May 18, 2026Redis launches Iris platformMemory architecture infrastructure
May 19, 2026MCP tunnels and self-hosted sandboxes launchPrivate enterprise access enablement

🔺 Scout Intel: What Others Missed

Confidence: high | Novelty Score: 82/100

While coverage focuses on individual announcements—MCP tunnels as a security feature, Cognition’s valuation as market validation, observational memory as a benchmark result—the deeper story is the convergence of infrastructure, capital, and architecture around enterprise-ready agent ecosystems.

MCP tunnels eliminate the deployment friction point, not just the security risk. The real economic impact isn’t reduced attack surface (though that matters); it’s removing the 2-4 week firewall change request process from deployment timelines. Enterprises can now deploy agents against internal databases in hours, not months. This compresses the sales cycle for agent vendors and the value realization timeline for buyers.

The $10B+ club reveals four distinct business models, not one market. Cursor (developer IDE, 25x ARR), Cognition (autonomous engineer, 342x implied multiple on current ARR), Sierra (enterprise CX, BPO model), Lovable (vibe coding, 15M DAU) are not competing—they’re segmenting. Enterprise buyers should evaluate against their specific workflow, not compare valuations across models. Cognition’s 73x ARR growth in 9 months signals autonomous coding demand, but the implied valuation multiple reflects growth expectations, not sustainable revenue.

Observational memory’s 10x cost reduction is the enterprise-relevant metric, not the benchmark score. LongMemEval scores (94.87% vs 80.05%) matter for research benchmarks, but CFOs will care about the 10x token cost reduction. At scale, this is the difference between economically viable long-running agents and projects killed at the budget review. Hybrid retrieval intent tripling in Q1 2026 confirms enterprises hit RAG scale limits and are actively seeking alternatives.

Protocol competition is over; dual adoption is the answer. MCP and A2A under the same Linux Foundation governance means enterprises can stop waiting for a winner. The ROI timeline difference (MCP 2-3 months, A2A 6-12 months) reflects architectural scope: MCP connects agents to tools (immediate automation value), A2A coordinates agents (requires workflow redesign). Plan for both.

Key Implication: Enterprises with security governance already in place can deploy MCP tunnels immediately and show ROI in weeks. Those without should prioritize governance documentation (one of the four attributes of the 12% hitting production) before infrastructure investment.

Outlook & Predictions

Near-term (0-6 months)

  • MCP tunnel adoption accelerates: Enterprises with existing perimeter security teams will deploy tunnels within weeks. Expect firewall change request volumes to drop as agent deployments shift to tunnel architecture. Confidence: high.

  • Observational memory pilots replace RAG for long-context workloads: Teams running into RAG scale limits will test observational memory on long-running agent projects. Cost reduction data (10x) will drive quick POC-to-production decisions. Confidence: medium.

  • Valuation multiples compress: The implied 342x multiple on Cognition’s current ARR is unsustainable. Expect valuation expectations to normalize toward the 25x-33x range demonstrated by Cursor and Lovable. Confidence: medium.

Medium-term (6-18 months)

  • Protocol governance matures under AAIF: With both MCP and A2A under Linux Foundation, expect enterprise-grade certification programs, security audits, and compliance frameworks. Confidence: high.

  • Memory architecture consolidation: Redis Iris, Mastra observational memory, and context management platforms will converge on standardized APIs. The “context layer” becomes a distinct infrastructure category. Confidence: medium.

  • Enterprise agent production rate doubles: From current 12% to 25%+ as infrastructure (MCP tunnels), memory (observational architectures), and protocols (MCP/A2A) mature simultaneously. Confidence: medium.

Long-term (18+ months)

  • Agent operating systems emerge: The current tooling landscape (MCP for connectivity, A2A for coordination, observational memory for context) will consolidate into integrated agent OS platforms. Confidence: low.

  • Valuation tier separation deepens: The $10B+ club will split: Cursor/Cognition (developer/eng tools) vs. Sierra (enterprise CX) vs. Lovable (consumer vibe coding). Cross-segment competition will be minimal. Confidence: medium.

Key Trigger to Watch

MCP tunnel deployment velocity: Track how quickly enterprises move from pilot to production with MCP tunnels. If deployment timelines remain compressed (weeks not months), expect rapid acceleration in enterprise agent adoption in Q3-Q4 2026. If security teams create new approval processes that restore month-long delays, adoption will plateau.

Sources

AI Agent Weekly Intel (W33): MCP Tunnels, Valuation Club, Memory Architecture Shift

MCP tunnels enable private enterprise access without firewall changes through outbound-only gateways. The $10B+ valuation club reveals distinct models: Cursor 25x ARR, Cognition 73x growth, Sierra enterprise CX. Observational memory cuts costs 10x vs RAG.

AgentScout · ·
#mcp-tunnels #ai-agents #memory-architecture #enterprise-deployment #valuation #protocols
Analyzing Data Nodes...
SIG_CONF:CALCULATING
Verified Sources

TL;DR

Three converging threads define this week’s AI agent ecosystem: MCP tunnels fundamentally restructure enterprise deployment economics by eliminating inbound firewall requirements through outbound-only gateway architecture; a tiered $10B+ valuation club emerges with Cursor ($50B, 25x ARR), Cognition ($25B, 73x growth), Sierra ($15.8B), and Lovable ($6.6B) showing distinct business model patterns; and observational memory architecture displaces RAG on long-context benchmarks (94.87% vs 80.05%) while cutting token costs 10x.

Key Facts

  • Who: Anthropic (MCP tunnels), Cognition ($25B valuation target), Mastra (observational memory), Linux Foundation (A2A governance)
  • What: Infrastructure security maturation, valuation acceleration, memory architecture evolution, protocol competition
  • When: May 19-24, 2026 (MCP tunnels launch), April-June 2026 (valuation announcements), Q1 2026 (hybrid retrieval tripling)
  • Impact: 150+ organizations on A2A protocol, 12% enterprise production rate, 33.3% hybrid retrieval adoption intent

Executive Summary

The AI agent ecosystem underwent three structural shifts this week that will shape enterprise deployment decisions for the next 18 months. First, Anthropic’s launch of MCP tunnels at ‘Code with Claude’ London on May 19, 2026, introduces an outbound-only gateway architecture that eliminates the need for inbound firewall rules—agents access private databases and internal APIs through a lightweight gateway running inside the enterprise network that establishes encrypted mTLS connections to Anthropic infrastructure. This architecture removes the single largest barrier to enterprise adoption: security team approvals for perimeter changes.

Second, the $10B+ valuation club crystallized into four distinct business model patterns: Cursor’s developer IDE model ($50B valuation, 25x ARR multiple, 67% Fortune 500 penetration), Cognition’s autonomous engineer approach ($25B target, 73x ARR growth from $1M to $73M in nine months), Sierra’s enterprise customer experience platform ($15.8B post-money after $950M raise), and Lovable’s vibe coding for non-technical users ($6.6B valuation, 15 million daily active users). Each targets a different segment with fundamentally different unit economics.

Third, observational memory architecture emerged as a viable alternative to RAG for long-running agent workloads. Mastra’s open-source implementation achieved 94.87% on the LongMemEval benchmark with GPT-5-mini (84.23% with GPT-4o), surpassing RAG baseline (80.05%) while reducing token costs by a factor of 10. Redis entered this space with the Iris platform launch on May 18, 2026, offering Context Retriever and Agent Memory components that run 99% of data on SSD at one-tenth the cost of in-memory storage. Enterprise adoption intent for hybrid retrieval tripled from 10.3% to 33.3% in Q1 2026, signaling that standalone RAG programs are hitting scale walls.

The protocol layer also saw consolidation. A2A (Agent-to-Agent) reached 150+ organizations in production at its one-year mark, while MCP was donated to the Linux Foundation’s Agentic AI Foundation (AAIF) with Anthropic, Block, and OpenAI as co-founders. These protocols are complementary: MCP standardizes tool and data connectivity (showing ROI in 2-3 months), while A2A enables agent-to-agent coordination (taking 6-12 months for full value). Enterprises should plan for dual-protocol deployments.

Background & Context

The AI agent infrastructure landscape evolved rapidly from experimental prototypes in 2024 to production deployments in 2026. Three forces drove this acceleration: the emergence of standardized protocols (MCP in December 2025, A2A in May 2026), the validation of autonomous coding agents as enterprise tools (Cognition’s Devin reaching $73M ARR), and the maturation of security frameworks (Five Eyes alliance publishing guidance on agent governance in April 2026).

Enterprise adoption patterns shifted from isolated pilot projects to integrated workflow deployments. The PwC-Anthropic alliance expansion on May 14, 2026—training and certifying 30,000 professionals on Claude—signals that professional services firms now view agentic technology not as experimental but as a core capability requiring systematic workforce development. G2 research indicates enterprises expect 10-25% of workflows to be managed by agents within months.

The memory architecture problem remained unsolved. RAG (Retrieval-Augmented Generation) programs hit scalability limits as context windows expanded but retrieval accuracy degraded. Hybrid retrieval adoption intent tripled in Q1 2026, reflecting a search for alternatives. Observational memory—using background agents to compress and maintain stable context windows—emerged as a viable solution for long-running agents, though not for dynamic knowledge base scenarios.

Infrastructure Security Maturation: MCP Tunnels

Outbound-Only Gateway Architecture

MCP tunnels represent a fundamental shift in how AI agents access private enterprise resources. The architecture operates on a simple principle: instead of requiring enterprises to open inbound firewall ports for agent access, a lightweight gateway runs inside the enterprise network and establishes an outbound encrypted mTLS connection to Anthropic’s tunnel control plane.

The security implications are substantial. Traditional agent deployments required enterprises to expose internal APIs and databases to the public internet, creating attack surfaces that security teams resisted. MCP tunnels eliminate this exposure by keeping all private resources behind the enterprise perimeter—the agent never holds credentials, and the gateway authenticates to the tunnel control plane.

Deployment Pattern:

ComponentLocationFunction
MCP GatewayInside enterprise networkOpens outbound mTLS connection to Anthropic
Tunnel Control PlaneAnthropic infrastructureRoutes agent requests through gateway
Private MCP ServersBehind enterprise firewallInternal databases, APIs accessed via tunnel
Agent SessionAnthropic infrastructureNo direct access to private resources

Source: Claude API Docs - MCP Tunnels Security

Enterprise Deployment Economics

The economic impact of removing firewall change requests from the deployment timeline is difficult to overstate. In most enterprises, firewall rule changes require approval from security, networking, and compliance teams—a process measured in weeks, not days. MCP tunnels reduce this timeline from “wait for approval” to “deploy gateway and configure outbound proxy.”

Thomson Reuters’ Legal MCP integration provides a concrete example. The deployment provides access to 1.9 billion Westlaw and Practical Law documents, 1.4 billion KeyCite validity signals, and a patent-pending citation ledger—all through MCP without requiring perimeter changes. For enterprises with similar internal knowledge bases, MCP tunnels represent immediate deployment enablement.

Governance Framework Requirements:

The Five Eyes alliance guidance published April 30, 2026, identifies autonomous agent behavior as a new enterprise attack surface requiring visibility and real-time controls. MCP tunnels address credential control at the network boundary, but enterprises must still implement:

  • Least-privilege role assignments for agent actions
  • RBAC (Role-Based Access Control) integration
  • Context limits to prevent runaway token consumption
  • Policy enforcement points at the gateway level
  • Continuous monitoring of agent-initiated actions

Source: Five Eyes ‘Careful Adoption of Agentic AI Services’

What This Means for Enterprises

For Security Teams: MCP tunnels reduce perimeter risk but shift security responsibility to identity governance and monitoring. The outbound-only model aligns with zero-trust principles—verify every request, never trust the network.

For Infrastructure Teams: Deployment velocity increases by an order of magnitude. A gateway can be deployed in hours, not weeks. The trade-off is dependency on Anthropic’s tunnel control plane—enterprises should negotiate SLAs and failover procedures.

For Procurement: The infrastructure conversation shifts from “we need firewall changes” to “we need gateway capacity planning.” Budget allocation moves from network perimeter expansion to internal gateway scaling.

Valuation Acceleration: The $10B+ Club

The AI agent market now has four distinct valuation tiers with fundamentally different business models. Understanding these models is essential for enterprise buyers evaluating build vs. buy decisions and for investors assessing market positioning.

Cursor: Developer IDE ($50B Valuation)

Cursor’s $50B valuation on $2B ARR (as of February 2026) represents a 25x multiple—exceptionally high for a developer tools company. The valuation reflects 67% Fortune 500 penetration and projected $6B ARR by year-end. Cursor 3, launched April 2026, introduced parallel agents and multi-session interfaces, directly competing with Claude Code’s desktop redesign.

Business Model: Enterprise subscriptions and developer seats. Revenue driver shifted from individual developers to enterprise contracts.

Key Metric: 67% Fortune 500 penetration indicates enterprise adoption validation.

Source: TechCrunch - Cursor $50B Valuation

Cognition: Autonomous Software Engineer ($25B Valuation Target)

Cognition’s ARR trajectory is unprecedented: $1M (September 2024) to $73M (June 2025)—73x growth in nine months. The $25B valuation target, if achieved, would more than double the $10.2B valuation from September 2025. Devin, the autonomous software engineer, can be pointed at a Jira ticket and will write code, run tests, and open PRs with minimal human intervention.

Business Model: Subscription to autonomous coding agent. Customers pay for engineering capacity without hiring.

Key Metrics: $1M to $73M ARR in 9 months, total net burn under $20M across company history (capital-efficient growth), Windsurf acquisition ($250M) for IDE capabilities.

Enterprise Partnerships: Infosys (first large digital services firm to deploy Devin at scale), Cognizant (AI-driven development collaboration).

Source: Cognition Official Blog

Sierra: Enterprise Customer Experience ($15.8B Valuation)

Sierra raised $950M at $15.8B post-money valuation—a 60% increase from the previous $10B round. Founded by Bret Taylor (OpenAI Chairman, former Salesforce Co-CEO), Sierra positions as “global standard” for AI-powered customer experiences. Unlike narrow support chatbots, Sierra agents handle full customer lifecycle: support, sales, account management.

Business Model: “Productized BPO” (Business Process Outsourcing). Sierra doesn’t just sell software—it takes on customer service operations.

Key Metric: $950M raise led by Tiger Global and GV, total capital now exceeds $1B.

Source: Sierra Official Blog

Lovable: Vibe Coding for Non-Technical Users ($6.6B Valuation)

Lovable targets the largest addressable market: non-technical users who want to build software without learning to code. The platform has 15 million daily active users and 200,000 new projects created daily. Projected $200M ARR for 2026.

Business Model: Subscriptions for non-technical creators. Lower per-seat price but massive volume.

Key Metrics: 15M DAU, 200K daily projects, $200M+ projected ARR (33x multiple at $6.6B).

Valuation Club Comparison

CompanyValuationARRARR MultipleTarget SegmentKey Differentiator
Cursor$50B$2B (Feb 2026)25xDevelopersIDE integration, Fortune 500 penetration
Cognition$25B$73M (Jun 2025)342x (implied)Engineering teamsAutonomous agent, 73x growth
Sierra$15.8BNot disclosedN/AEnterprise CXFull lifecycle, BPO model
Lovable$6.6B$200M+ (2026 proj)33xNon-technical usersVibe coding, 15M DAU

Strategic Insight: The highest multiples (Cursor 25x, Lovable 33x) attach to platforms with demonstrated enterprise penetration or massive user bases. Cognition’s implied 342x multiple on current ARR reflects growth expectations, not current revenue. Sierra’s valuation reflects founder pedigree and enterprise CX market size, not ARR multiple.

Memory Architecture Evolution: Observational Memory vs. RAG

The RAG paradigm—retrieve relevant documents, inject into context window, generate response—is hitting fundamental limits as agent workloads become more sophisticated. Three problems dominate: retrieval accuracy degrades as knowledge bases expand, context windows consume tokens at unsustainable rates, and long-running agents accumulate context that overwhelms retrieval systems.

Observational memory addresses these problems through a different architecture: background agents continuously compress and curate context, maintaining stable windows (~30k tokens) that provide relevant information without retrieval overhead.

Benchmark Results: LongMemEval

Mastra’s open-source observational memory implementation achieved 94.87% on LongMemEval with GPT-5-mini, surpassing RAG baseline (80.05%) by 14.82 percentage points. With GPT-4o, observational memory scored 84.23%—still 4.18 points above RAG.

Architecture Components:

ComponentFunctionToken Impact
Observer AgentMonitors all interactions, extracts salient informationAdds minimal overhead
Reflector AgentGarbage collects irrelevant observations when threshold (~40k tokens) reachedMaintains ~30k stable window
Three-Tier RepresentationProgressively compresses informationPrioritizes critical context

Cost Impact: 10x reduction in token costs. Stable context windows enable aggressive caching, while RAG requires fresh retrieval queries for each interaction.

Source: Mastra Research - Observational Memory

When to Choose: Observational Memory vs. RAG vs. Hybrid

ArchitectureBest ForCost ProfileAccuracy (LongMemEval)
Observational MemoryLong-running agents with stable context needs10x reduction94.87% (GPT-5-mini)
RAG BaselineDynamic knowledge bases, changing informationStandard retrieval costs80.05%
Hybrid RetrievalEnterprise RAG hitting scale limitsContext-dependentNot benchmarked

Enterprise Adoption Signal: Hybrid retrieval adoption intent tripled from 10.3% to 33.3% in Q1 2026 (VentureBeat VB Pulse data), indicating enterprises are moving away from standalone RAG toward context architectures.

Source: VentureBeat - Context Architecture Replacing RAG

Redis Iris Platform Entry

Redis launched the Iris platform on May 18, 2026, with five components targeting agent memory and context management:

  1. Context Retriever: Makes external data navigable by agents
  2. Agent Memory: Persistent context storage
  3. Redis Data Integration: Connects enterprise data sources
  4. LangCache: Caching layer for LLM responses
  5. Redis Search: Semantic search over context

Cost Efficiency: Redis Flex runs 99% of data on SSD at one-tenth the cost of in-memory storage, making large-scale context storage economically viable.

Source: Redis Official Blog

Protocol Competition: A2A vs. MCP

The agent interoperability layer is crystallizing around two protocols that serve different purposes. Understanding when to adopt each is critical for enterprise architecture decisions.

Protocol Comparison

DimensionMCPA2A
PurposeTool/data connectivityAgent-to-agent coordination
GovernanceLinux Foundation AAIFLinux Foundation AAIF
AdoptionUniversal (Anthropic standard)150+ organizations in production
ROI Timeline2-3 months6-12 months
Enterprise SupportAnthropic, Block, OpenAI, Google, Microsoft, AWSGoogle, Microsoft, AWS, Salesforce, SAP, ServiceNow, IBM
Use Case”How do agents interact with tools?""How do agents work together?”

Source: TrueFoundry - MCP vs A2A, Linux Foundation - A2A Milestone

Governance Consolidation

Both protocols now fall under the Linux Foundation’s Agentic AI Foundation (AAIF), donated by Anthropic in December 2025 with Anthropic, Block, and OpenAI as co-founders. Google, Microsoft, AWS, Cloudflare, and Bloomberg support the foundation.

This consolidation signals that the protocol wars are ending—enterprises can safely adopt both without fear of fragmentation. The protocols are complementary:

  • MCP: Adopt for every data source and tool you want agents to access. Shows immediate ROI through automation acceleration.
  • A2A: Adopt when operating multiple AI systems that need to coordinate workflows. Requires more architectural planning.

Security Note: Neither protocol was designed with adversarial input as a first-class concern. Enterprises must layer additional security controls—input validation, output filtering, behavior monitoring—on top of protocol-level trust.

Enterprise Deployment Patterns

Production Metrics: What Distinguishes Successful Deployments

Enterprise data from Digital Applied shows 12% of AI agent projects reach production. Successful deployments share four attributes:

  1. Pre-deployment infrastructure investment: Security, identity, monitoring built before agent deployment
  2. Governance documentation: Clear policies on agent authority, escalation procedures, audit trails
  3. Baseline metrics: Measurement frameworks established before deployment to demonstrate ROI
  4. Dedicated business ownership: Agents have accountable owners, not just technical custodians

Source: Digital Applied - Enterprise Adoption Data

Claude Code Parallel Agent Workflows

The Claude Code desktop redesign (April 14, 2026) introduced production-grade parallel agent management:

Key Features:

  • Multi-session sidebar: Manage active and recent sessions
  • Git worktree isolation: Each session scoped to .claude/worktrees/ directory, zero context bleed
  • Three view modes: Verbose, Normal, Summary
  • Cloud-based Routines: Scheduled automations running on Anthropic infrastructure

Agent View: Command center for steering multiple agents across parallel tasks. Emerging data shows 3-5 parallel sessions typical for complex projects.

Source: Claude Official Blog - Desktop Redesign

PwC-Anthropic Alliance: Enterprise Transformation Blueprint

The expanded alliance (May 14, 2026) establishes patterns for enterprise-wide agent deployment:

Three Focus Areas:

  1. Agentic technology build: Helping engineering teams build AI agent tools for clients
  2. AI-native deal-making: Deploying AI across M&A and deal processes
  3. Enterprise function reinvention: Reinventing operating models with AI

Scale: Joint Center of Excellence, 30,000 PwC professionals to be trained and certified on Claude. Claude-native finance business group launched.

ROI Patterns: Enterprises report 3x-5x ROI within 12-18 months for aggressive deployments. The critical success factor is workflow reimagination—redesigning processes around agent capabilities—not simply adding agents to existing workflows.

Source: PwC Official - Anthropic Alliance Expansion

Key Data Points

MetricValueSourceDate
Cognition ARR Growth73x ($1M to $73M in 9 months)Cognition Official BlogJun 2025
Observational Memory LongMemEval Score94.87% (GPT-5-mini)Mastra ResearchMay 2026
Observational Memory vs RAG+14.82 pts (94.87% vs 80.05%)Mastra ResearchMay 2026
Token Cost Reduction10xVentureBeatMay 2026
Hybrid Retrieval Adoption Intent10.3% → 33.3% (Q1 2026)VentureBeat VB PulseMar 2026
A2A Protocol Adoption150+ organizations in productionLinux FoundationMay 2026
Cursor Valuation$50B (25x ARR)TechCrunchApr 2026
Cursor Fortune 500 Penetration67%TechCrunchApr 2026
Sierra Valuation$15.8B post-moneyTechCrunchMay 2026
Lovable Daily Active Users15 millionPanto AI AnalysisApr 2026
Enterprise Production Rate12%Digital AppliedQ1 2026
PwC Claude Certification Target30,000 professionalsPwC OfficialMay 2026
Redis Flex Cost Efficiency99% data on SSD at 1/10th costRedis OfficialMay 2026
MCP Tunnels Launch’Code with Claude’ LondonAnthropic OfficialMay 19, 2026

Timeline: AI Agent Ecosystem (December 2025 - May 2026)

DateEventSignificance
Dec 2025Anthropic donates MCP to Linux Foundation AAIFInfrastructure standardization milestone
Dec 2025Cognition acquires Windsurf ($250M)AI IDE market consolidation
Apr 2026Cursor 3 launch with parallel agentsCompetitive response to Claude Code
Apr 14, 2026Claude Code desktop redesignProduction-grade parallel workflows
Apr 17, 2026Cursor $50B valuation talks$10B+ club tier formation
Apr 23, 2026Cognition $25B valuation talksAutonomous engineer market validation
Apr 30, 2026Five Eyes publishes agent governance guidanceGovernment-level security framework
May 1, 2026A2A protocol: 150+ organizations in productionProtocol adoption milestone
May 4, 2026Sierra raises $950M at $15.8BEnterprise CX market maturation
May 14, 2026PwC-Anthropic alliance expansionEnterprise deployment acceleration
May 18, 2026Redis launches Iris platformMemory architecture infrastructure
May 19, 2026MCP tunnels and self-hosted sandboxes launchPrivate enterprise access enablement

🔺 Scout Intel: What Others Missed

Confidence: high | Novelty Score: 82/100

While coverage focuses on individual announcements—MCP tunnels as a security feature, Cognition’s valuation as market validation, observational memory as a benchmark result—the deeper story is the convergence of infrastructure, capital, and architecture around enterprise-ready agent ecosystems.

MCP tunnels eliminate the deployment friction point, not just the security risk. The real economic impact isn’t reduced attack surface (though that matters); it’s removing the 2-4 week firewall change request process from deployment timelines. Enterprises can now deploy agents against internal databases in hours, not months. This compresses the sales cycle for agent vendors and the value realization timeline for buyers.

The $10B+ club reveals four distinct business models, not one market. Cursor (developer IDE, 25x ARR), Cognition (autonomous engineer, 342x implied multiple on current ARR), Sierra (enterprise CX, BPO model), Lovable (vibe coding, 15M DAU) are not competing—they’re segmenting. Enterprise buyers should evaluate against their specific workflow, not compare valuations across models. Cognition’s 73x ARR growth in 9 months signals autonomous coding demand, but the implied valuation multiple reflects growth expectations, not sustainable revenue.

Observational memory’s 10x cost reduction is the enterprise-relevant metric, not the benchmark score. LongMemEval scores (94.87% vs 80.05%) matter for research benchmarks, but CFOs will care about the 10x token cost reduction. At scale, this is the difference between economically viable long-running agents and projects killed at the budget review. Hybrid retrieval intent tripling in Q1 2026 confirms enterprises hit RAG scale limits and are actively seeking alternatives.

Protocol competition is over; dual adoption is the answer. MCP and A2A under the same Linux Foundation governance means enterprises can stop waiting for a winner. The ROI timeline difference (MCP 2-3 months, A2A 6-12 months) reflects architectural scope: MCP connects agents to tools (immediate automation value), A2A coordinates agents (requires workflow redesign). Plan for both.

Key Implication: Enterprises with security governance already in place can deploy MCP tunnels immediately and show ROI in weeks. Those without should prioritize governance documentation (one of the four attributes of the 12% hitting production) before infrastructure investment.

Outlook & Predictions

Near-term (0-6 months)

  • MCP tunnel adoption accelerates: Enterprises with existing perimeter security teams will deploy tunnels within weeks. Expect firewall change request volumes to drop as agent deployments shift to tunnel architecture. Confidence: high.

  • Observational memory pilots replace RAG for long-context workloads: Teams running into RAG scale limits will test observational memory on long-running agent projects. Cost reduction data (10x) will drive quick POC-to-production decisions. Confidence: medium.

  • Valuation multiples compress: The implied 342x multiple on Cognition’s current ARR is unsustainable. Expect valuation expectations to normalize toward the 25x-33x range demonstrated by Cursor and Lovable. Confidence: medium.

Medium-term (6-18 months)

  • Protocol governance matures under AAIF: With both MCP and A2A under Linux Foundation, expect enterprise-grade certification programs, security audits, and compliance frameworks. Confidence: high.

  • Memory architecture consolidation: Redis Iris, Mastra observational memory, and context management platforms will converge on standardized APIs. The “context layer” becomes a distinct infrastructure category. Confidence: medium.

  • Enterprise agent production rate doubles: From current 12% to 25%+ as infrastructure (MCP tunnels), memory (observational architectures), and protocols (MCP/A2A) mature simultaneously. Confidence: medium.

Long-term (18+ months)

  • Agent operating systems emerge: The current tooling landscape (MCP for connectivity, A2A for coordination, observational memory for context) will consolidate into integrated agent OS platforms. Confidence: low.

  • Valuation tier separation deepens: The $10B+ club will split: Cursor/Cognition (developer/eng tools) vs. Sierra (enterprise CX) vs. Lovable (consumer vibe coding). Cross-segment competition will be minimal. Confidence: medium.

Key Trigger to Watch

MCP tunnel deployment velocity: Track how quickly enterprises move from pilot to production with MCP tunnels. If deployment timelines remain compressed (weeks not months), expect rapid acceleration in enterprise agent adoption in Q3-Q4 2026. If security teams create new approval processes that restore month-long delays, adoption will plateau.

Sources

q72lk1iaacavu3tgy214xk████4c9vor606bekr40x3fuh9cuxr98hk12f░░░wuw8muf8yq7lxszar3qhjpbj961a8nq8████r5o2o4l0w5ua9r6q8en9s06uv421d6████3w320m2zfvxu4e98lgkxqlq2t2o1ryy████bm0mgyhwzdkib52is2ez9ftaraajtc5░░░k2726lpzrjjzx9x1v6k66y8q8jad598████t6m1eozuk1lxmes2gxaelum6zc07ekg░░░jcprariisekbmn4xkdy4ll4u9cxekug░░░ppfhfae1llc7pc2fg551tkr3njsgahja████4vq4b5te383o6qte9288ymwk00tzsa8g░░░6vjw0h9zundqezy7wq4kwb7sqiwounxza░░░rhf06qjnrum202q6orih84z5h8e0g41wr░░░dq5n86gv6squugv4ubihxie2n1cdw8pfc░░░ufmbt8rp7li2xddamac1rjega40h2b████mxeeudj2dy87ywx97gshapoj6rmfenfm████o7wzofzgwwl694y5u9vpcj95it469mne████jpyobxd4x0lzvifzl9knjb0w472hi27i████y2ka4ikdpftdxfh6wutk46rbil3jotfg████mucunmgc6qejtrra8cs4emdepx3ee692████o8yy23767wak0ty8fr2yih9ptx8lcnx8░░░cz6oq3vl7gn1ux9i9ljxxq7m8bby6xm2h████sl31ww786dntz2n20p1ie0ylfzlgfnf3d░░░vhlj69ru42okjrj7egvkts3b9plz8e6yh████gj18q0fhpqvk8epqj1ccsexbiiiwru8████m6r3c2391vrskqi3xtrw6axyivj24z65t████em7yei8x5ocb2arjl4rzudhbrf8g45dl░░░qujejvgglmjar2xbukdfmbo2yrc6ie28s░░░cdng6wql6bfxfi9i0ltm4rfv0lbyngzen░░░er0fhnrdx34ciw7r985iyeb89qouls8h░░░gk28ayq5sv8k8x3o8uwfy5oo9qpz0sap░░░uspwmwcyj1n9kezytqcpstrntp3i0lnnm░░░uk8y3n00wksn275od0suha43f2ofzuot░░░u1qcdm7771d5czisyfzior73j5kekx8░░░6kpslb5bzunfd4wl1hlb8bz8m1plzus9████9k846bk8m9gizkoafq8lffq7ldovfjvr████rrrsefz86smhj0ddqr9hom7m5ncyo5m░░░b3iid1x6qwkdgu3j4uut3wlh1r3ercvag████9fou0mk5fiby2tzkj0a1wa37hdgolrzt████4kyagd8qvzo6vbnviwhkbwsies5ascjg░░░oj6wf0zkbj90zbj6nuoqd1ldbgdtgmp1░░░01mrrftg4wrzu7ux2uu243ph7esjn4295b░░░8385b1ajy55phj5cuhdt2go87jyjouqdo░░░frj6r5f5pbmfullpk1dnpmaex53951gyr████yc88iajbhnij2aqjuic8li33l2oe13l████f8ed28iqh4g4yg2m7bls17xfwzixb7ieo████ier7q3eh3ap0m915sup50omttz2bmh7f2g████vwldp2xsxpda3xj91c8x8h6opwnj2jxp░░░gsamxvhy2lvwzmhblqmrceb1yg9q3kgwr░░░u6m52qghdsexwas0jj2x6y8l5kjeoln░░░wcdf3a8bx6