aiGalen Guan

Hermes Agent v0.18.0 Deep Dive: The 208K-Star Self-Evolving AI Agent — MoA, Verification Loops, and the P0/P1 Engineering Miracle

Hermes Agent v0.18.0 Deep Dive: The 208K-Star Self-Evolving AI Agent

On July 22, 2025, Nous Research released Hermes Agent on GitHub. One year later: 208,669 stars, 38,006 forks, MIT license, Python.

This is one of the fastest-growing AI agent projects on GitHub. Its positioning is uniquely distinctive:

The agent that grows with you. The only agent with a built-in learning loop — it creates skills from experience, improves them during use, nudges itself to persist knowledge, searches its own past conversations, and builds a deepening model of who you are across sessions.

What Hermes Agent Is

Hermes Agent is a self-evolving AI agent platform. It's not another "ChatGPT wrapper" — it's a complete agent operating system:

  • A real terminal interface: Full TUI with multiline editing, slash-command autocomplete, conversation history, interrupt-and-redirect, streaming tool output
  • Lives where you do: Telegram, Discord, Slack, WhatsApp, Signal, and CLI — all from a single gateway process
  • A closed learning loop: Agent-curated memory with periodic nudges, autonomous skill creation after complex tasks, skills self-improve during use, FTS5 session search with LLM summarization for cross-session recall
  • Scheduled automations: Built-in cron scheduler with delivery to any platform — all in natural language, running unattended
  • Delegates and parallelizes: Spawn isolated subagents for parallel workstreams, Python scripts that call tools via RPC
  • Runs anywhere: Six terminal backends — local, Docker, SSH, Singularity, Modal, and Daytona

Core Architecture: The Self-Evolving Learning Loop

Hermes Agent's most differentiating capability is its closed learning loop:

Use → Create Skills → Improve Skills → Persist Memory → Cross-Session Recall → Smarter Use

This loop is driven by four subsystems:

1. Skills System

Hermes's skill system is its "procedural memory" — reusable workflows and expertise:

  • Autonomous creation: After completing complex tasks (5+ tool calls), the agent automatically proposes creating a skill
  • Self-improvement during use: When a skill is found outdated or incomplete during use, the agent auto-patches it
  • agentskills.io standard: Compatible with the open skill standard, shareable across agents
  • 116+ built-in skills: Covering coding, research, creative, data science, DevOps, and more

v0.18.0's new /learn command simplifies skill creation to the extreme — point at a directory, URL, or just-completed workflow, and the agent auto-distills it into a reusable skill.

2. Memory System

Hermes's memory system is its "declarative memory" — durable facts about you and the world:

  • Dual-track memory: USER.md (who you are) + MEMORY.md (the agent's notes)
  • Proactive nudges: The agent periodically reminds itself to persist important information
  • Cross-session search: FTS5 full-text search + LLM summarization for cross-session context recall
  • Honcho dialectic user modeling: Builds a deepening model of who you are

v0.18.0's new /journey command and desktop memory graph (a playable radial timeline) make memory visible — you can see what your agent knows, watch it grow, and prune what's wrong.

3. Delegation System

Hermes's delegation system supports background fan-out — spawning multiple subagents to work in parallel:

  • Isolated contexts: Each subagent has its own conversation, terminal session, and toolset
  • Background execution: The parent session is never blocked; results come back as a single consolidated turn when all subagents finish
  • Up to 3 concurrent: Configurable concurrency limit

v0.18.0 upgraded delegation from "blocking wait" to "background fan-out" — kick off "research these five competitors in parallel" or "audit these three modules," then carry on with something else.

4. Cron System

Hermes's cron system supports natural language scheduled tasks:

  • Natural language scheduling: "every morning at 6am," "every 30 minutes"
  • Multi-platform delivery: Results can be delivered to Telegram, Discord, local files, etc.
  • Skill-driven: Cron jobs can load skills to execute complex workflows
  • Script support: Python/Bash scripts as data collectors

v0.18.0: The Judgment Release — Core Breakthroughs

Released July 1, 2026, v0.18.0 is a major milestone for Hermes Agent. ~1,720 commits, 998 merged PRs, 949 issues closed, 370+ community contributors.

Breakthrough 1: P0/P1 Zero — An Engineering Miracle

This is the release headline. The team cleared every P0 and P1 issue/PR across the entire repo in 12 days:

Priority Issues Closed PRs Merged
P0 (critical) 3 8
P1 (high) 493 188
Total 496 196

~692 highest-priority items resolved in twelve days. The team commits to keeping P0/P1 at zero from here forward.

Breakthrough 2: Mixture-of-Agents as a First-Class Citizen

MoA went from "a mode you toggle" to "a model you pick":

  • Every named MoA preset appears as a selectable model under a moa provider
  • Sits alongside Claude, GPT, and Grok in every model picker (CLI, TUI, desktop, gateway)
  • Pick "my-council" the same way you'd pick any model — Hermes routes your prompt through that ensemble automatically

More critically, transparency: each reference model's full output now renders as its own labeled block — you can read what GPT-5 thought, what Claude thought, and what Grok thought, before the aggregator synthesizes them into one answer. And that final answer now streams to you live instead of appearing all at once after a long silence.

Breakthrough 3: Verification Loop — "Done" Means Proven, Not Claimed

Hermes now records verification evidence for coding work and can decide it's finished by actually running your project's checks, not by asserting success:

  • /goal gained completion contracts: you state what "done" looks like, and the standing-goal loop judges completion against that evidence
  • pre_verify hook for wiring in custom checks
  • One-time migration that tunes the defaults sensibly

The difference between "I think I fixed it" and "the tests pass, here's proof."

Breakthrough 4: /learn and /journey — Self-Evolution Visualized

  • /learn <anything>: Point at a directory, URL, or just-completed workflow, and the agent auto-distills a reusable skill, written to your CONTRIBUTING.md standards
  • /journey: A playable learning timeline in CLI and TUI, showing the memories and skills Hermes has accumulated — editable or deletable right from the view
  • Desktop memory graph: A top-down, playable radial timeline of memories and skills

Your agent's memory stops being a black box.

Breakthrough 5: Desktop Coding Projects

The desktop app gained real coding Projects:

  • Sidebar of your codebases
  • Coding rail
  • Review pane
  • Git worktree management
  • Agent-facing project tools
  • project → repo → lane model

The desktop transformed from a "chat window" into a "coding cockpit."

Breakthrough 6: Scale-to-Zero Deployment

The gateway can now go dormant when idle and quiesce cleanly before a restart, migration, or auto-update — without dropping in-flight conversations. A hosted or relay-only Hermes can scale to zero when nobody's talking to it and wake back up on demand.

Version Evolution: From Surface to Judgment

Version Date Theme Core Changes
v0.16.0 6/5 The Surface Release Native desktop app (macOS/Linux/Windows), 100 PRs, 159 commits in one week
v0.17.0 6/19 The Reach Release iMessage + WhatsApp channels, deeper tool integration, team features
v0.18.0 7/1 The Judgment Release P0/P1 zero, MoA first-class, verification loop, /learn + /journey, background fan-out

Three releases, three themes: Surface (desktop) → Reach (coverage) → Judgment (quality). This is textbook version planning — lay the surface first, expand coverage, then polish quality.

Tech Stack Highlights

Component Technology Highlight
Language Python 3.11+ Full-stack Python, uv package management
Terminal Full TUI Multiline editing, slash-command autocomplete, streaming tool output
Desktop Native app macOS/Linux/Windows, built in one week across 100 PRs
Gateway Multi-platform messaging Telegram/Discord/Slack/WhatsApp/Signal/iMessage
Backends 6 types Local/Docker/SSH/Singularity/Modal (serverless)/Daytona
Tools 40+ Terminal, file, web, browser, computer use, code execution, etc.
Skills 116+ built-in agentskills.io standard, autonomous creation + self-improvement
Memory FTS5 + LLM Dual-track memory, Honcho user modeling, cross-session search
Models Any Nous Portal/OpenRouter/OpenAI/Anthropic/local — no lock-in

Comparison with Similar Projects

Dimension Hermes Agent OpenClaw Claude Code Codex Cursor
Stars 208K 381K
Positioning Self-evolving agent platform Personal AI assistant Terminal coding agent Terminal coding agent IDE agent
Learning loop ✅ Core
Skill system ✅ Auto-create+improve
Memory system ✅ Dual-track+cross-session
Multi-platform messaging ✅ 7 platforms
Desktop app ✅ Native ✅ IDE
MoA ✅ First-class
Verification loop ✅ Completion contracts
Cron scheduling ✅ Built-in
Delegation/parallel ✅ Background fan-out
Serverless ✅ Modal/Daytona
License MIT MIT Proprietary Proprietary Proprietary

Hermes Agent is unique on three dimensions: learning loop, MoA, and verification loop. It's not another coding assistant — it's an agent operating system that grows.

Multi-Dimensional Scoring

Dimension Score Notes
Architecture 9/10 Self-evolving learning loop is paradigm-level innovation; skills+memory+delegation+cron four-subsystem synergy
Engineering Quality 9/10 P0/P1 zero is an engineering miracle; 692 high-priority items in 12 days, 370+ contributors
User Experience 9/10 CLI/TUI/desktop/messaging four-interface unification; /learn and /journey make self-evolution visible
Platform Coverage 9/10 7 messaging platforms + 6 terminal backends + 3 desktop OSes — nearly ubiquitous
Innovation 10/10 MoA first-class, verification loop, self-evolving skills — multiple world firsts
Community Vitality 10/10 208K stars, 370+ contributors per release, 25,418 open issues indicate extreme activity
Maturity 7/10 One year old, rapid iteration, 25,418 open issues indicate significant work remaining

Overall: 9.0/10. Hermes Agent is currently the most complete and innovative open-source AI agent platform. Its self-evolving learning loop is a paradigm-level innovation in the agent space — the agent doesn't just execute tasks, it grows through execution.

Limitations and Risks

  1. 25,418 open issues: While P0/P1 are at zero, the total issue count is enormous, indicating the project is still rapidly expanding with edge cases under-covered.
  2. Complexity: Skill system, memory system, delegation system, cron system — extremely high concept density, steep learning curve for new users.
  3. Python ecosystem dependency: Full-stack Python, less friendly to non-Python developers.
  4. Resource consumption: A full-featured agent (MoA + multiple tools + gateway) requires significant compute resources.
  5. Rapid iteration risk: Frequent version updates (major release every two weeks), backward compatibility may be affected.

Conclusion

Hermes Agent is a gift from Nous Research to the AI agent ecosystem. It proves a thesis: the best AI agent isn't driven by the most powerful model, but by the one that learns best.

Its design philosophy is worth deep reflection for every AI system builder:

  1. Agents should grow through use: Skills auto-create and self-improve, memory accumulates across sessions — the agent is not a static tool, it's a dynamic companion
  2. "Done" needs proof, not claims: The verification loop transforms the agent from "I think I'm done" to "the tests pass, here's proof"
  3. Transparency matters more than intelligence: MoA shows every model's reasoning, /journey shows the agent's learning history — users need to see the agent's thinking
  4. Ubiquitous, but not binding: 7 messaging platforms, 6 backends, any model — the agent should adapt to the user, not the other way around

Our recommendation: if you're looking for an AI agent platform that grows, Hermes Agent is currently the best open-source choice. Start with hermes setup --portal, experience the self-evolution capabilities of /learn and /journey, then gradually explore MoA, cron, and the delegation system.

For AI researchers: Hermes Agent's self-evolving learning loop is one of the most important architectural innovations in the agent space — the design of the skill system, memory system, and verification loop deserves deep study.

References