Agenthatch – Compile any skill into a standalone Python agent Agenthatch launches a tool that compiles SKILL.md files into standalone Python agents, addressing isolation, validation, and scaling issues in multi-skill AI workflows. The deterministic pipeline parses markdown, runs six AI harnesses, and generates a runnable package with typed tools and MCP integration, reducing token waste and enabling unlimited skill scaling. Turn any SKILL.md into a standalone, runnable AI Agent. SKILL.md promised a lot. Write a markdown file, tell your agent what to do, and it works. In practice, anyone who has used more than three skills across Claude Code, Codex CLI, or OpenClaw knows the friction: | Pain point | What actually happens | |---|---| No isolation | Skills leak into each other. A file-organizer skill and a git-ops skill share the same context window. The agent confuses instructions meant for one with the other. | Reference book, not operating manual | Agents treat SKILL.md as a loose suggestion, not a contract. Given a long skill, the model skim-reads it. It picks the parts that seem relevant and ignores the rest. | Token waste | Every SKILL.md lives in the system prompt. Add 5 skills at 3KB each and you just burned 15KB of context before the conversation even starts. On long-running tasks this compounds fast. | No validation | A typo in a tool name, a missing parameter, an ambiguous instruction. The agent won't catch any of it until runtime, and by then the conversation is 20 turns deep. | Scale decays | Skills work at 1–3. At 10+ they become unmanageable. No dependency graph, no conflict detection, no way to tell which skill overrides which. | The core issue isn't the format. It's that SKILL.md is prompt engineering , not software engineering . You're asking an LLM to interpret human prose at runtime, every time, with no compilation, no type checking, no contract. agenthatch treats a SKILL.md as source code — not a prompt. It compiles it through a deterministic pipeline into a standalone Python agent that you can import, ship, and run anywhere. SKILL.md → Parse → 6-Harness LLM Pipeline → Code Generation → Runnable Agent input Phase 1 Phase 2: AI inference Phase 3: Jinja2 output The result is a self-contained Python package with its own pyproject.toml , a CLI entry point, typed tool definitions, MCP integration, and a runtime configuration. It's not a wrapper around your skill — it is the skill, compiled into code. Install pip install agenthatch Initialize with your LLM provider agenthatch init Add a SKILL.md agenthatch skills add ./my-skill/SKILL.md Hatch it into an agent agenthatch hatch my-skill Run it agenthatch run my-skill Three steps from markdown to running agent. The hatched agent lives in your skillhouse and can be re-run anytime. | SKILL.md raw | agenthatch hatched | | |---|---|---| Execution | Interpreted at runtime by LLM | Compiled into standalone Python package | Isolation | All skills share one context window | Each agent has its own runtime, tools, and config | Validation | None. Typos and ambiguities caught at runtime. | Schema-validated AHSSPEC before code generation | Token cost | Full skill body in system prompt every turn | ~150 bytes of runtime config | Tool definitions | Prose descriptions, LLM guesses how to call | Type-annotated Python functions with JSON Schema | MCP | Manual wiring per agent | Auto-detected, auto-configured | Determinism | LLM interprets differently each time | Same SKILL.md → same AHSSPEC structure low-temp inference | Multi-skill scaling | Degrades past 3–5 skills | Unlimited. Each agent is a separate process. | Debugging | Read the LLM's chain-of-thought and pray | Standard Python debugging, logging, tests | agenthatch runs a 3-phase pipeline with 6 AI harnesses working in parallel: The SKILL.md is parsed for frontmatter, body, and directory files. No AI involved. A pure file-system operation. The output is a ContextPack with zero semantic transformation. Six specialized AI harnesses process the skill, each with its own persona and temperature profile: | Harness | Role | Temp | |---|---|---| A — Identity | Extract name, version, description from frontmatter | 0.1 | B — Intent | Infer trigger phrases and user intents | 0.5 | C — Interface | Design tool signatures, parameters, and return types | 0.5 | D — Base | Detect runtime base class and instruction structure | 0.3 | E — Assembly | Cross-validate all harness outputs, produce AHSSPEC | 0.2 | F — MCP | Detect and configure MCP server connections | 0.3 | Each harness runs an Analyze → Infer → Self-Validate → Correct loop with up to 2 internal retries. Harness E cross-validates the other five outputs and produces a unified AHSSPEC Agent Hatch Standard Specification . Jinja2 templates render the AHSSPEC into a complete Python agent package: hatched-agent/ ├── pyproject.toml pip-installable package ├── runtime.toml LLM provider, model, API keys ├── README.md Generated usage docs ├── agenthatch.yaml AHSSPEC manifest └── src/{package name}/ ├── init .py ├── agent.py Agent class extends AHCoreAgent ├── tools.py Type-annotated tool implementations └── references.py AI-extracted structured data Generated agents use the PlanLayer state machine — a 6-state planning engine that runs STARTING → PLANNING → EXECUTING → VERIFYING → REPLANNING → DONE. It adapts mid-task: merges completed steps, branches on failure, and degrades gracefully when tools time out. Click to expand: the full pipeline in detail Sets up ~/.agenthatch/ with your LLM provider configuration. Supports OpenAI, DeepSeek, Anthropic, and any OpenAI-compatible endpoint. The config file is TOML. Readable, versionable, easy to share. Copies the SKILL.md and its directory into the skillhouse index. The skillhouse tracks every skill you've added, its hatch status, and where its generated agent lives. The full pipeline runs: Phase 1 deterministic : Parse SKILL.md → ContextPack Phase 2 AI : 6 harnesses → HarnessOutput → Assembly → AHSSPEC Phase 3 Jinja2 : AHSSPEC → agent package Flags: --no-generate — skip Phase 3, review the AHSSPEC first --force — overwrite existing hatched agent --dry-run — preview without writing files Launches the hatched agent in interactive TUI mode. The agent loads its runtime config, connects to its LLM provider, and starts a conversation loop with tool calling, context compaction, and PlanLayer-driven execution. | Command | What it does | |---|---| agenthatch init | Initialize config and provider setup | agenthatch skills add