Agent Ecosystem

April 2026 research briefing

AI Agent Skills Ecosystem shows rapid 10x growth in five months

This page is designed to answer the simple question first: where is the real momentum? The short answer is that developer adoption is accelerating across agent skills, protocols are becoming shared infrastructure, and security has become the deciding constraint for trust and scale.

Executive summary

The market is maturing in public

OpenClaw created the early distribution wave, Hermes proved that learning loops can be more controlled, and Claude Code is showing how enterprise workflows absorb these patterns. Read the top line first, then use the lower sections as evidence.

OpenClaw Stars
345K+
In 5 months
ClawHub Skills
44K+
7.7x in 2 months
MCP Downloads
97M
Monthly SDK downloads
Hermes Stars
64K+
In 3 months, Zero CVEs

Daily Skill Submission Velocity

10x Growth
Mid-Jan 202650/day
Early Feb 2026500+/day

Growth accelerated before the OpenClaw acquisition and ClawHavoc incident, indicating genuine developer adoption rather than media-driven spikes.

OpenClaw

Gateway-first, TypeScript/Node.js, unbounded memory, marketplace model with 44K+ skills.

345K Stars 9+ CVEs <1.2s latency

Hermes Agent

Learning-first, Python-native, 5-stage loop, bounded memory (~1,300 tokens), auto-generated skills.

64K Stars Zero CVEs 40% faster repeat

Claude Code

Enterprise tier, 4% of GitHub commits, projected 20% by end 2026, compliance-focused.

Enterprise 4% Commits Pentagon Risk
CHAPTER 1-3

Framework Comparison

This section compares the main players in plain terms. OpenClaw is strongest in distribution, Hermes is strongest in structured learning, and Claude Code is strongest where enterprise controls matter most.

GitHub Star Trajectory

Framework Stars

OpenClaw surpassed React's 10-year milestone in 5 months

Performance & Security

Framework Comparison

Nous Research benchmarks, April 2026

Detailed Framework Matrix

Dimension OpenClaw Hermes Claude Code
Center of GravityGateway (communications)Agent loop (learning)Editor integration
Skill PhilosophyHuman-authored, marketplace (44K+)Auto-generated from experience (118 bundled)Skills + shell commands
Memory ModelUnbounded Markdown files (LanceDB)Bounded, curated (~1,300 tokens)Session-based
Latency<1.2s median~30% slower routingVariable
Tokens/Turn~1,800~8,000+ (with trajectory)Variable
Security (Apr 2026)9+ CVEs, 135K+ exposedZero agent-specific CVEsEnterprise-managed
Repeat Task SpeedBaseline40% fasterN/A
LanguageTypeScript/Node.jsPythonProprietary
CHAPTER 1

Ecosystem Scale & Growth

These numbers show whether the category is becoming real infrastructure or just a burst of attention. The pattern here points to durable adoption: users, instances, downloads, and revenue are all moving together.

+925% MoM
38M
Monthly website visitors
Active
3.2M
Monthly active users
82 countries
500K+
Running instances
180 startups
$320K+
Monthly ecosystem revenue
5,800+
MCP
Servers & 300+ clients
Viral
34K
Peak stars in 48 hours

Key Growth Metrics Table

MetricValuePeriodContext
OpenClaw Stars345,000+Nov 2025 – Mar 2026Passed React's 10-year trajectory
ClawHub Skills44,000+Feb – Apr 20267.7x increase in ~8 weeks
Daily Submissions50 → 500+/dayMid-Jan – Early Feb10x acceleration
skills.sh (top skill)20,000 in 6 hoursJan 20, 2026Launch day velocity
MCP Downloads97 million/moMar 2026970x in 16 months
Hermes Stars64,000+Feb – Apr 202630% migration from OpenClaw
CHAPTER 2 & 4

Skills Taxonomy

This chapter explains what people are actually building. The most active skill groups are practical rather than decorative: automation, research, development, and task execution dominate the mix.

ClawHub Skills by Official Category

Community-Curated Functional Groups

AI & LLM Orchestration280+
Search & Research250+
DevOps & Cloud210+
Web & Frontend200+
Productivity & Tasks130+
Browser & Automation322
Coding Agents & IDEs500+

Top categories are overwhelmingly action-oriented, not chat-oriented.

Top 15 Skills by Downloads

#SkillCategoryDownloadsFunction
1Capability EvolverAI/ML35,000+AI self-evolution engine
2GOG (Google Workspace)Productivity33,800+Gmail, Calendar, Drive, Docs
3Self-Improving AgentAI/ML32,000+Cross-session learning
4OntologyAI/ML27,600+Typed knowledge graph
5SummarizeProductivity26,100+URLs, PDFs, audio, video
6GitHub IntegrationDevelopment24,800+Full GitHub workflow
7Felo SearchResearch145,000+AI-synthesized search
8NotionProductivity13,900+Read/write pages & DBs
9Nano Banana ProCreative13,400+Image generation (Gemini)
10API GatewayDevelopment13,000+Managed OAuth to 100+ APIs
11ObsidianProductivity12,400+Local vault integration
12Agent BrowserWeb11,800+Rust headless browser
13OpenAI WhisperMedia11,500+Local speech-to-text
14McPorterDevelopment11,100+MCP server management
15Tavily SearchWeb/AI9,000+AI-optimized web search

Coding & Development

GitHub (24.8K), DevOps (393 skills), Browser automation, "Night shift" patterns

Key: Docker, K8s, AWS, Terraform, Playwright

Content Creation & Media

164 image/video skills, audio/music, presentations (54 design systems)

Key: PixVerse, Suno, ElevenLabs, Felo Slides

Business Automation

GOG (#2 downloaded, 33.8K), SEO Research (highest ROI), CRM

Key: Linear, Notion, HubSpot, Autobound

AI/ML & Research

MLOps (17 bundled in Hermes), arXiv, Search (250+ skills), Red teaming

Key: vLLM, TRL, Hugging Face, Tavily, Bright Data
CHAPTER 6

Security Landscape: The ClawHavoc Crisis

This is the cost of speed without enough trust controls. The security story matters because it changes how marketplaces, enterprise deployment, and protocol governance need to be designed going forward.

1,184
Malicious Skills Discovered
Across 12 publisher accounts; hightower6eu responsible for 57%
36.82%
Skills with Security Flaws
Snyk ToxicSkills audit: 1,467 of 3,984 scanned skills
135K+
Exposed Instances
93% without proper auth; 12,812 RCE-exploitable

CVE Severity Distribution

CVE Severity

ClawHavoc Attack Timeline

Jan 24
Clawdbot released; mass adoption begins
Jan 27
First malicious skill: polymarket-trading-bot v1.0.1
Jan 31
Mass deployment: 386 malicious skills in 1 day (354 from hightower6eu)
Feb 1
Koi Security publishes "ClawHavoc" report; public disclosure
Feb 7
OpenClaw partners with VirusTotal; automated scanning begins

Snyk ToxicSkills Audit Results

MetricCountRateAssessment
Total scanned3,984Largest public corpus
Any security flaw1,46736.82%>1 in 3 skills affected
Critical issues53413.4%>1 in 8 critically vulnerable
Confirmed malicious761.9%Human-verified payloads
Plaintext credentials2837.1%API keys in LLM context/logs
Prompt injection + malware91%of maliciousDual-vector evasion

OWASP Agentic Skills Top 10 (AST10)

AST01
Malicious Skills
Critical — ClawHavoc evidence
AST02
Supply Chain Compromise
Critical — ClawHub collapse
AST03
Over-Privileged Skills
High — 280+ credential-leaking
AST04
Insecure Metadata
High — Fake brand impersonation
AST05
Unsafe Deserialization
High — YAML payload delivery
AST06
Weak Isolation
High — Host-mode execution
AST07
Update Drift
Medium — ClawJacked CVE
AST08
Poor Scanning
Medium — NL evades detection
AST09
No Governance
Medium — 53K exposed, no SOC
AST10
Cross-Platform Reuse
Medium — Multi-registry porting
CHAPTER 5

Infrastructure & Protocols

Here we look at the pipes underneath the ecosystem. What we need to have is a shared way for tools, agents, and companies to connect safely (protocol layer), and this is where MCP, A2A, and ACP begin to matter.

MCP
97-110M
Monthly SDK downloads
Agent-to-tool | JSON-RPC 2.0 | Linux Foundation (AAIF)
A2A
150+
Partner organizations
Agent-to-agent | Google | Linux Foundation
ACP
3 Variants
Agent Communication Protocol
Cisco AGNTCY | IBM BeeAI | Zed Editor

Protocol Comparison Matrix

FeatureMCPA2AACP (IBM)ACP (Cisco)ACP (Zed)
Primary FocusAgent-to-ToolAgent-to-AgentAgent-to-AgentAgent-to-AgentEditor-to-Agent
TransportJSON-RPC 2.0, HTTP, SSEHTTP(S), JSON-RPC 2.0, SSEHTTP RESTful APIRESTful APIsJSON-RPC 2.0 over stdio
DiscoveryApplication-integratedAgent Cards (.well-known)Embedded metadataAgent Directory + OASFEditor plugin registry
GovernanceLinux Foundation (AAIF)Linux FoundationIBM / OpenCisco AGNTCYZed + JetBrains
CreatedNov 2024 (Anthropic)Apr 2025 (Google)May 2025 (IBM)2024 (Cisco)2025 (Zed)
Key Metric97-110M/mo downloads150+ partnersBeeAI integrationAGNTCY coalitionv0.11.0, 2.3K stars

Distribution Channels

ClawHub
Centralized registry, VirusTotal + Gemini scanning, auto-hide on reports
skills.sh (Vercel)
Cross-platform CLI, 20K+ installs, .well-known discovery
GitHub CLI
gh skill CLI, content-addressed change detection, version pinning
Private/Enterprise
Workspace skills, managed directories, custom GitHub taps

Knowledge Management Integrations

Obsidian
Three-layer architecture (Intake → Retrieval → Integration), MCP servers, personalized RAG
Notion
Official MCP (18 tools, 4K+ stars), used as control plane for 18-agent deployments
Logseq
Graph-based note-taking, AI plugins with RAG
Memory Systems
Supermemory, Mem0, agentmemory, Letta/MemGPT, Cognee, Zep/Graphiti

Distribution Channel Trust Matrix

ChannelMechanismTrust ModelSecurity FeaturesKey Limitation
ClawHubCentralized registryCommunity + auto scanVirusTotal + Gemini Code InsightNo pre-publication review; no signing
skills.shDecentralized GitHub reposGitHub identity + telemetrySHA-256 via well-known endpointNo security scanning of content
GitHub CLIGit-native distributionGitHub auth + immutable releasesContent-addressed, version pinningNo cryptographic signing
Private/EnterpriseCustom repos + workspace dirsOrganization-controlledDetermined by org policyNo standardized RBAC
CHAPTER 7

Growth Patterns & Traction

Growth is not uniform. Some segments are expanding because they solve immediate work problems, while others are growing because security and governance suddenly became urgent after public failures.

Top Skill Categories by Count

Top Categories

Framework Growth Trajectory

Growth Trajectory

Security Skills

Post-ClawHavoc boom in scanning, auditing, and vulnerability detection tools. Fastest-growing segment.

AI/LLM Orchestration

280+ skills — model routing, prompt engineering, autonomous agents. Largest category on ClawHub.

DevOps & Cloud

210+ skills reflecting the "operator not chatbot" pattern. Direct embodiment of autonomous ops.

GEO

Generative Engine Optimization — $7.3B market by 2031. Optimizing for AI search citations.

Developer Migration

30% active developer migration from OpenClaw to Hermes seeking "maintenance fatigue" relief.

Night Shift Economy

Claude Code: 4% of GitHub commits, projected 20%. Jules: 140K+ autonomous code improvements.

Enterprise Adoption Signals

SignalMetricConfidence
Gartner: enterprise apps agentic by end 202640%, up from <5% in 2025High
OpenClaw running instances500K+ across 82 countriesHigh
OpenClaw monthly active users3.2 millionHigh
OpenClaw monthly website visitors38 million (925% MoM growth)High
MCP monthly SDK downloads97 million (970x in 16mo)High
Claude Code: % of GitHub commits4% → 20% projected by end 2026High
Agentic AI project failure risk (Gartner)40% abandoned by 2027 without governanceHigh
CHAPTER 8

Emerging Sectors & Future Outlook

This is the frontier view. We need to have reliable coordination between software agents, payment rails, local devices, and governance systems (agent networks, machine payments, edge inference, compliance tooling) for the next phase to work.

Emerging Sectors Overview

SectorKey Metric (2026)MaturityKey PlayersStrategic Implication
Agent social networks1.5M API tokens exposed; $93M market cap crashNascent / high riskMoltbook, Permiso, WizAgent-to-agent communication requires security-first architecture
Agent-to-human labor markets518K+ workers; 32.7% programmatic originEarly / scalingRentAHuman.aiPhysical-world attack surface extends beyond digital boundaries
Agent-native payments165M+ transactions; 480K+ transacting agentsGrowthx402 Foundation, Coinbase, StripeMachine-to-machine payments becoming infrastructure
Local/edge deploymentApple Silicon <200ms; RPi 300–500msEmergingApple, NVIDIA, QualcommSovereign deployment now technically viable
Agent security tooling36.82% flaw rate; 1,184 malicious skillsRapid growthOWASP, Snyk, CiscoSecurity tooling is the fastest-growing sub-sector
Guardian/governance agents10–15% projected market share by 2030NascentAccelirate, GartnerCompliance automation as a service
Protocol consolidationMCP 97M downloads; A2A 150+ orgsStandardizingAnthropic, Google, IBM, Linux FoundationInteroperability becoming table stakes

The Agent-to-Agent Economy

Moltbook
Agent social network. 1.5M registered agents, $93M market cap crash, 500+ prompt injection attacks.
RentAHuman.ai
518K+ human workers available for agent-hired tasks. $25 median per task. 32.7% programmatic origin.
x402 Protocol
165M+ transactions, $50M+ volume, 480K+ transacting agents. USDC on Base, <$0.001 fees.

Sovereign & Local AI Deployment

Apple Silicon (M-series)<200ms
NVIDIA Jetson~250ms
Qualcomm Snapdragon~350ms
Raspberry Pi / ARM Cortex300-500ms

Agent inference no longer requires data center GPUs. A $100 Raspberry Pi can run capable agents.

Security-Learning Paradox

The most efficient distribution model (centralized marketplace) is also the most vulnerable. The most secure model (autonomous generation) sacrifices network effects.

Markdown's Unpatchable Attack Surface

Natural language instructions evade both code scanners and AI safety filters. 91% of malicious skills combined prompt injection with traditional malware.

Protocol Stack Consolidation

MCP (vertical) + A2A (horizontal) + ACP (enterprise governance) = the new TCP/IP for agents. Interoperability commoditizes framework differences.

Knowledge Management = Real Moat

Obsidian vaults, Notion control planes, and Google Workspace histories create higher switching costs than the frameworks themselves.

CHAPTER 9

Strategic Playbook

This final section turns the research into action. Instead of treating every framework as a winner, it separates where each one is strongest and what decisions different audiences should make next.

For Developers

  • OpenClaw for breadth/connectivity, Hermes for depth/learning, Claude Code for enterprise
  • Treat skills as untrusted code — sandbox, scan with mcp-scan, review SKILL.md source
  • Target underserved verticals: health, education, manufacturing

For Enterprises

  • Governance first: RBAC, audit trails, least privilege before wide deployment
  • Hermes' bounded memory for compliance-sensitive environments (finance, healthcare, gov)
  • Vendor risk assessments on all AI agent vendors — including open-source

For Investors

  • Security infrastructure is the highest-conviction sector post-ClawHavoc
  • Protocol layers beat application layers: MCP's 97M downloads vs any single framework
  • The "night shift economy": time arbitrage through autonomous agents as a new value category

Paid Service & Tooling Opportunities

Security Scanning-as-a-Service

Semantic skill scanners that understand instruction intent. The gap is structurally larger than npm's 2015 gap. Estimated 10-15% of total agent infrastructure spending by 2030.

Night Shift Orchestration

Scheduling, monitoring, and QA for autonomous overnight work. Solopreneurs report 70-90% time savings. Infrastructure rents comparable to cloud computing's capex-to-opex shift.

Vertical Skill Marketplaces

HIPAA-compliant medical skills, educational assessment generators, manufacturing QC inspectors. Less competition, established procurement budgets.

Governance & Compliance Agents

"Guardian Agents" projected 10-15% market share by 2030. EU AI Act logging, GDPR compliance, audit trail automation. Gartner: 30% higher trust, 25% better compliance.

Knowledge Management Integration

The real moat. Higher switching costs than frameworks. Build specialized connectors for Obsidian, Notion, Confluence with three-layer RAG architectures.

Agent-to-Agent Payment Rails

x402-style micropayment infrastructure. 165M+ transactions prove demand. Build escrow, dispute resolution, and KYT/OFAC compliance layers.