Agent Harness Kit

Agent Harness Kit is a local execution-governance layer for coding agents. It gives Codex and Claude Code durable project context, adaptive workflow selection, dependency-aware execution, ownership leases, proportional verification, and inspectable completion.

Install

uv tool install agent-harness-kit-cli

For an existing installation, run uv tool upgrade agent-harness-kit-cli and verify agent-harness --version; repeating uv tool install does not upgrade an already installed tool.

Or install from the same PyPI release with pipx or plain pip, including from the integrated terminal in VS Code:

pipx install agent-harness-kit-cli
python -m pip install agent-harness-kit-cli

On Windows, py -m pip install agent-harness-kit-cli is also supported. Prefer a virtual environment with plain pip.

Open a terminal inside a project and run:

agent-harness install

The recommended core profile is installed by default. Use --dry-run to preview or --profile core-learning to include optional, consented project-learning support.

The installer creates a contained agent-harness-kit/ directory plus root AGENTS.md and CLAUDE.md entrypoints. Missing root files are created; existing instructions are preserved outside one managed bridge block. Open a new agent context at the project root after installation, then use agent-harness doctor to verify the entrypoints.

What it does

• four execution lanes (direct-trivial, vibe, graph-only, full-harness) independent from none|light|full assurance;
• evidence-first resume: probe current repository/runtime/check state, then read durable artifacts only to fill gaps;
• adaptive first-run project-shape discovery: reuse approved evidence, resolve architecture and folder organization before planning, and ask about optional coding conventions only when they are not already established;
• executable preflight for declared paths, scripts, environment names, commands, validator, browser/sandbox requirements, and worker capacity;
• PENDING.md for human decisions and macro gaps;
• TASK-GRAPH.md for dependencies, leases, progress, and execution;
• compact same-context specs and consumer-driven handoff/review artifacts;
• planned units targeting 15–30 active minutes, with justified exceptions;
• focused-first verification that climbs through workspace, integration, checkpoint, and delivery only as needed;
• scoped read_set, exclusive write_set, related impact_set, and source provenance to avoid unnecessary repository rescans;
• maximum-cardinality collision-free scheduling for hosts with proven parallel capacity;
• atomic compare-and-swap graph transitions, hash-chained events, append-only run metrics, and lighter-lane suggestions;
• bounded native Codex dispatch packets and fresh-context review when the host provides those capabilities;
• hackathon mode for compressed discovery and demo-first MVP delivery;
• a validator covering source, packaged profiles, and installed-host behavior.

The CLI performs deterministic installation, inspection, routing, preflight, scheduling, state transitions, metrics, and dispatch-packet operations. Capable agent hosts perform the actual coding and agent creation. The Kit has no unattended background daemon and does not lock operating-system files, merge branches, or deploy software on its own.

Documentation, source, audio overview, and examples are available in the GitHub repository.