MeshHarmes · agent harness

The rules of agent work
written in code.

An agent harness and project infrastructure for general Python work, where the game rules are written in code, not in a prompt: an entry gate that decides whether work can proceed, a backlog with acceptance criteria, memory outside the context window, and a permission gate for irreversible actions.

github.com/cacelass/MeshHarmes ← back to projects
01
problem What it solves
An AI generates code much faster than a human reviews it. MeshHarmes is the environment that puts the reins on: a repository to start projects where an AI agent does part of the work, with a gate, a verifiable backlog and a permission system, agnostic of the AI provider, in pure Python with zero dependencies.
Agents
24 Python agents, one responsibility each
Purity
agents/ is stdlib-only, working before anything is installed
Gate
init.sh decides if the project is healthy enough to work on
02
decisions Engineering decisions
→ The gate is code, not convention: harness finish refuses to close a feature unless the gate passes with real command evidence.
→ One resource, one owner: only the harness agent writes the backlog; the implementer is the only one that touches product code.
→ Memory outside the context window: progress lives in files (harness/progress/), searchable via RAG after each close.
→ A permission gate for irreversible actions: destructive actions are declared in agents/contracts.py and refuse to run without explicit authorisation.
→ Provider-agnostic: any coding agent that can run shell commands can use it; no vendor SDK.
→ Prompt-injection aware: external data is data, never instructions; irreversible actions still ask for confirmation.
Relationship MeshHarmes is the standalone evolution of the harness that ships inside dskit.