Agent AFK is an open-source AFK coding agent designed to run tasks while you are away. It plans, acts, verifies its own work, and notifies you when it finishes or needs input. Free, Apache-2.0, works with any model provider.

What does AFK mean for a coding agent?

AFK stands for Away From Keyboard. An AFK coding agent is designed to keep working after you walk away; it runs a task end to end, verifies the result, reaches a clear stopping point, and texts you when done.

How does Agent AFK notify me when it finishes?

Agent AFK sends a Telegram message when a run finishes, stalls, or needs your input. You can answer, redirect, or interrupt from your phone. Every run also writes a full trace to disk so you can see exactly what happened.

Which one learns faster — Hermes or Agent AFK?

Hermes learns faster, by design. Its background review loop creates and patches skills autonomously from session history, and the Honcho engine builds a deepening user model over time. Agent AFK does not have a learning loop — its 23 skills are hand-authored and static between releases. If self-improvement is a priority, Hermes is the stronger choice. AFK trades autonomous learning for a guaranteed terminal-state contract: consistent, verifiable delivery rather than improving generality.

Can I migrate from Hermes to Agent AFK?

Yes, with some manual work. Agent AFK operates on files and git repos independently of any prior agent. Hermes-generated skills carry over manually — there is no automatic migration path. Cross-session memory in AFK starts from zero and builds forward. They serve different workflows: Hermes is a general-purpose agent; AFK is a coding specialist.

Is Hermes only for Nous Research models?

No. Hermes supports OpenAI-compatible endpoints beyond the Nous family, though it performs best with Hermes models. Agent AFK supports Anthropic, any OpenAI-compatible provider, and local models (Ollama, LM Studio). Each AFK pipeline step can target a different provider independently — granularity Hermes does not currently expose.

Which is better for coding specifically?

Agent AFK. It is purpose-built for software engineering: 23 coding skills, DAG orchestration with worktree isolation, a full interactive REPL alongside the daemon, and a terminal-state contract with evidence. Hermes can code, but coding is one of many use cases. If your workflow is almost entirely software engineering, AFK’s specialization pays off.

Does Agent AFK have a learning loop?

No. Agent AFK’s 23 skills are hand-authored and updated between releases, not generated from session history. Cross-session memory stores facts and conventions the agent observes, but no background process creates or patches skill scripts autonomously. Hand-authored skills are predictable and auditable; for projects where reliability trumps novelty, AFK’s static skill set is an asset.

Can both tools be open source if Hermes is MIT and AFK is Apache-2.0?

Yes. Both licenses support commercial use, forking, and personal projects. MIT has minimal attribution requirements; Apache-2.0 adds explicit patent protection. Neither creates friction for typical developer use. Both repos are public, auditable, and forkable.

Comparison

Agent AFK vs. Hermes

Hermes, by Nous Research, is one of the most sophisticated open-source AI agents available. Its self-improving skill system creates and patches automation scripts from experience, and its Honcho-powered user model deepens across every session. Agent AFK does none of that — and does not try to. It is a CLI daemon built for one thing: unattended software engineering with verified, evidence-backed results. Two genuinely different tools optimized for different problems.

Self-improving assistant vs. verified coding daemon

Hermes’s design center is a general-purpose AI assistant that gets better the longer you use it. Its background review loop replays sessions, extracts reusable skills, and patches them when it finds a better approach. The Honcho dialectic engine builds a structured model of who you are and how you prefer to work — state that persists and sharpens across every conversation. Agent AFK’s design center is the developer who needs a task done while they’re away. Every turn ends in one of four terminal states — Done, Blocked, Asking, or Interrupted — and Done must cite an Evidence pointer: a file path, a commit SHA, or a test result. That contract is enforced in code, not prompt. Hermes has the deeper learning loop. Agent AFK has the deeper delivery guarantee. For software engineering specifically, AFK is purpose-built. Hermes is a general-purpose agent that happens to do some coding.

Feature comparison

Grounded in the actual implementations, not marketing copy.

Dimension	Hermes	Agent AFK
How a task ends	A response is returned. No structured terminal state — the agent does not distinguish between Done, Blocked, or needing clarification in a machine-readable way. You interpret the output.	One of four terminal states: Done, Blocked, Asking, or Interrupted. Done requires an Evidence pointer to a file, commit, or test result. Injected in code — no config can remove it.
Primary workflow	General-purpose assistant accessible across 20+ messaging platforms (Discord, Slack, Telegram, WhatsApp, and more), voice mode, and a web UI. Optimized for conversation-first interaction.	CLI daemon: you state a goal, the agent runs to a named terminal state, and the verdict arrives on your phone via Telegram. Absence is the design center. Pure developer focus.
Self-improvement / learning loop	Background review loop replays sessions, extracts reusable skill scripts, and patches them when a better approach is found. Skills accumulate and improve autonomously over time. Most sophisticated in open source.	23 hand-authored skills ship in the box. The skill library is static between releases — no background loop learns from your sessions. Reliability through explicit authorship, not autonomous generation.
Memory / user modeling	Honcho dialectic engine builds a persistent, structured user model across sessions. 7+ external memory backends (Mem0, Zep, Letta, and more). Background review synthesizes long-term context automatically.	SQLite with FTS5 full-text search. Durable fact archive across sessions, queryable and injected into the system prompt for high model adherence. Simpler than Honcho but purpose-built for developer context.
Multi-agent orchestration	Agent-to-agent delegation is supported. No explicit DAG primitive or worktree isolation. Parallel hypothesis investigation is not a first-class workflow pattern.	Full DAG orchestration via the `compose` tool: parallel layers, fail-fast semantics, and git worktree isolation per subagent. /diagnose forks parallel hypotheses into isolated worktrees natively.
Developer tooling	7 terminal backends (Docker, SSH, Modal, Daytona, Singularity, Vercel Sandbox, local). Code execution, browser automation, and file operations. Broad tooling surface for a general-purpose agent.	23 coding-specific skills: /mint (spec → plan → build → verify), /diagnose, /ship, /review, /fix-pr, /ground-state, and more. Auto-routed by task shape. Full body loads only when a skill fires.
Messaging platforms	20+ platforms natively: Discord, Slack, Telegram, WhatsApp, Signal, Matrix, Messenger, Teams, IRC, and more. Voice mode included. Built for meeting users wherever they are.	Telegram only — for async push notifications of task verdicts. The Telegram integration is intentional: it is the receipt, not the channel. Agent AFK is a CLI tool, not a chatbot.
Scheduling	No native cron scheduler. Tasks are triggered by conversation or platform event. Automated recurring workflows require external tooling or custom integration.	Built-in cron scheduler with a persistent daemon. Schedule any skill or prompt to run unattended. Results land on Telegram. Full programmatic access — pipe it into whatever automation you already have.
Model providers	Supports Nous Research’s own Hermes model family plus OpenAI-compatible endpoints. Works best with Hermes models; broader provider support varies by capability.	Anthropic + any OpenAI-compatible endpoint + local models (Ollama, LM Studio, etc.). Each pipeline step can target a different provider or cost tier independently.
License	MIT licensed. Fully open source.	Apache-2.0. Fully open source. No feature gates.

Where Agent AFK pulls ahead

Genuine architectural differences, not marketing claims.

Terminal-state contract with evidence

Every Agent AFK turn ends in Done, Blocked, Asking, or Interrupted — one of four, every time. Done must carry an Evidence line: a file path, commit SHA, or test result. Hermes returns a response; you decide whether the task is done. AFK’s contract is enforced in code, not prompt — it is the same guarantee whether you run a one-liner or a 12-skill DAG.

DAG orchestration with worktree isolation

/diagnose forks parallel hypotheses into isolated git worktrees. The `compose` primitive executes a dependency graph with parallel layers and fail-fast semantics. Each subagent gets its own working tree so parallel branches cannot corrupt each other. Hermes supports agent delegation but has no equivalent DAG primitive or first-class worktree isolation.

Pure developer focus

23 hand-authored skills ship in the box, every one purpose-built for software engineering: /mint, /ship, /review, /fix-pr, /diagnose, /ground-state, and more. Auto-routed by task shape. Full skill bodies load only when a skill fires — no idle token cost. Hermes is a general-purpose assistant; coding is one of many use cases.

Full interactive REPL alongside the daemon

AFK is not only a background daemon. The interactive REPL ships a full TUI with 52 slash commands, ghost-text inline completions, tool-lanes that surface every tool call as it fires, and a thinking lane for visible reasoning. Work side-by-side in the terminal, then hand off to the daemon when you step away. Hermes offers a CLI TUI too, but AFK’s is tuned entirely for software engineering workflows.

TypeScript-native, same stack as your projects

Agent AFK is written in TypeScript and runs on Node. If your project is TypeScript, the agent’s own codebase is readable, forkable, and auditable without context-switching. Hermes is Python-based. For TypeScript teams, AFK’s stack reduces the surface area you need to reason about when something unexpected happens.

Where Hermes pulls ahead

Honest. Hermes has real strengths that Agent AFK does not match.

Self-improving skill system

Hermes’s background review loop is the most sophisticated in open source. After sessions complete, it replays them, identifies reusable patterns, and creates or patches skill scripts autonomously. Skills get better the more Hermes is used — without human authorship. Agent AFK’s skill library is static between releases.

Honcho dialectic user modeling

Honcho builds a persistent, structured model of who you are — your preferences, working patterns, communication style, and domain knowledge — and deepens it across every session. 7+ external memory backends support Mem0, Zep, Letta, and more. Agent AFK’s memory is a developer fact archive, not a personality model.

20+ messaging platforms and voice mode

Hermes meets users wherever they are: Discord, Slack, Telegram, WhatsApp, Signal, Matrix, Messenger, Teams, IRC, and a dozen more, plus voice mode. For teams that want an AI assistant embedded in their existing communication stack, Hermes is the clear choice. Agent AFK supports Telegram only, as a delivery channel for task verdicts.

Research lab backing from Nous Research

Hermes is built by Nous Research — the lab behind the Hermes, Nomos, and Psyche model families, with ~25,000 tests in the repo. Research depth influences the architecture in ways community projects rarely reach. Model and agent improvements are coordinated in one org.

7 terminal backends

Docker, SSH, Modal, Daytona, Singularity, Vercel Sandbox, and local execution. Hermes can run code in nearly any compute environment, making it well-suited for multi-environment workflows where execution context varies per task. Agent AFK runs locally or over SSH without built-in cloud sandbox abstractions.

Frequently asked questions

Which one learns faster — Hermes or Agent AFK?: Hermes learns faster, by design. Its background review loop creates and patches skills autonomously from session history, and the Honcho engine builds a deepening user model over time. Agent AFK does not have a learning loop — its 23 skills are hand-authored and static between releases. If self-improvement is a priority, Hermes is the stronger choice. AFK trades autonomous learning for a guaranteed terminal-state contract: consistent, verifiable delivery rather than improving generality.
Can I migrate from Hermes to Agent AFK?: Yes, with some manual work. Agent AFK operates on files and git repos independently of any prior agent. Hermes-generated skills carry over manually — there is no automatic migration path. Cross-session memory in AFK starts from zero and builds forward. They serve different workflows: Hermes is a general-purpose agent; AFK is a coding specialist.
Is Hermes only for Nous Research models?: No. Hermes supports OpenAI-compatible endpoints beyond the Nous family, though it performs best with Hermes models. Agent AFK supports Anthropic, any OpenAI-compatible provider, and local models (Ollama, LM Studio). Each AFK pipeline step can target a different provider independently — granularity Hermes does not currently expose.
Which is better for coding specifically?: Agent AFK. It is purpose-built for software engineering: 23 coding skills, DAG orchestration with worktree isolation, a full interactive REPL alongside the daemon, and a terminal-state contract with evidence. Hermes can code, but coding is one of many use cases. If your workflow is almost entirely software engineering, AFK’s specialization pays off.
Does Agent AFK have a learning loop?: No. Agent AFK’s 23 skills are hand-authored and updated between releases, not generated from session history. Cross-session memory stores facts and conventions the agent observes, but no background process creates or patches skill scripts autonomously. Hand-authored skills are predictable and auditable; for projects where reliability trumps novelty, AFK’s static skill set is an asset.
Can both tools be open source if Hermes is MIT and AFK is Apache-2.0?: Yes. Both licenses support commercial use, forking, and personal projects. MIT has minimal attribution requirements; Apache-2.0 adds explicit patent protection. Neither creates friction for typical developer use. Both repos are public, auditable, and forkable.

Compare other tools

See how Agent AFK compares across the landscape.

Try Agent AFK

Open source, Apache-2.0, bring your own model.

Install free Read the docs View on GitHub