Agent Harness Kit

Agent Harness Kit gives coding agents durable project context, separate human and technical pending views, dependency-aware execution, file ownership leases, bounded review, and inspectable completion.

Install

uv tool install agent-harness-kit-cli

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 and managed root AGENTS.md and CLAUDE.md bridge blocks without replacing existing project instructions.

What it coordinates

• approved project context before broad inspection;
• PENDING.md for human decisions and macro gaps;
• TASK-GRAPH.md for dependencies, leases, progress, and execution;
• automatic feature discovery plus optimized writing plans: non-trivial work becomes small executable task specs, while truly simple work stays inline;
• a direct-trivial fast path for localized color, spacing, typo, and static-label edits without SPEC, graph, TDD, or review ceremony;
• a deterministic-first request router with direct-trivial, vibe, graph-only, and full-harness lanes; AI classifies only genuine ambiguity;
• a zero-artifact vibe path for one low-risk local behavior change, with mandatory focused verification and automatic promotion when scope grows;
• a machine-readable agent-harness route "request" preflight that always returns one of the four lanes and safely marks unresolved ambiguity for AI refinement;
• test-driven code tasks with meaningful RED before production changes, minimal GREEN, and proportional regression evidence;
• scoped read_set, exclusive write_set, related impact_set, and source provenance to avoid unnecessary repository rescans;
• isolated workstreams and context-aware frontend/backend/integration routing;
• agent-driven parallel fan-out that fills proven capacity, waits for the first event, refills safely, and joins branches through explicit integration nodes;
• capability-tier model routing that applies confirmed model/reasoning overrides at Codex task or subagent dispatch when explicitly approved;
• executable native Codex agent dispatch with neutral-role selection, minimal context, fresh implementer/reviewer identities, and adapter-owned response evidence;
• automatic fresh-context review against the versioned task SPEC, with at most one focused re-review;
• hackathon mode for compressed discovery and demo-first MVP delivery;
• mandatory status with progress, blockers, next action, and inspectable paths.

The Kit is an artifact-driven scaffold followed by capable agents. It has no unattended background daemon. During an active orchestration session it can launch supported internal subtasks in parallel; it 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.