Agent Harness Kit, version zero point five point two.

Agent Harness Kit is an installable, platform-neutral scaffold for organizing software development with coding agents. It gives agents durable project context, clear task boundaries, focused verification, and a reliable way to continue work without repeatedly scanning the whole repository or asking for ceremonial approval.

Install the command-line tool with U V, then run agent harness install inside a project. The installer adds a contained kit plus safe root entrypoints. Codex enters through A G E N T S dot markdown, Claude Code through CLAUDE dot markdown, and both follow the same neutral rules and state.

On a new context window, resume request, or status request, the agent must read the approved project context first, PENDING dot M D second, and TASK GRAPH dot M D third. PENDING owns human decisions, human actions, and the macro view of unfinished project areas such as backend or authentication. TASK GRAPH owns technical order, dependencies, and execution. Every status must show stage, progress, blockers, next action, and inspectable paths. It may scan more broadly only when those artifacts reveal a concrete gap or the user explicitly requests an audit.

To control cost without silently reducing quality, tasks are routed by capability. Narrow mechanical work can use economical models, normal implementation uses balanced models, and frontier models are reserved for consequential architecture, security, difficult integration, or repeated failure. Model choice never grants extra permission.

Each task also carries an executable budget. Attempts, no-progress cycles, and context expansion are capped across the same goal, even when the model, agent, task, or session changes. When a ceiling is reached, the agent records evidence and replans instead of consuming tokens indefinitely.

It works for both new projects and mature repositories that already have Claude context, Codex instructions, agents, rules, or a custom harness. In those repositories, it does not overwrite anything immediately. It first inventories and freezes existing authorities. It then installs through staged coexistence in a separate namespace and requires human review before anything is removed or replaced. The README also links to the migration and coexistence playbooks and contracts.

The recommended Core profile covers normal delivery. Core Learning adds optional project-based learning, and Full also includes a separate pack for studying harness engineering. Installing learning support never activates observation or creates notes automatically. The agent first asks for the exact Markdown path, Obsidian folder, Notion target and connector, or another approved destination.

On first use, the agent presents the Kit and begins a short discovery before proposing technology or implementation. The welcome also explains that the user can choose standard delivery or hackathon mode. Hackathon mode compresses discovery, prioritizes one testable end-to-end slice, integrates early, and aims for a working M V P or demo before secondary polish.

After discovery, the user approves the project context and important decisions. The Kit creates human and macro pending work in PENDING, then technical order and dependencies in TASK GRAPH. Frontend, backend, data, infrastructure, integration, and learning can use separate contexts when the host supports them. Each active task receives bounded source context, an exclusive file lease, and a distinct reviewer.

When declared checks pass, the task is marked complete, reported, and the next ready task can start. Independent review runs as non-blocking assurance: one proportional review, and only for a real blocker, at most one focused re-review of its correction and related regressions. Hackathon work uses a light review by default. There is never a third review loop.

The README now leads with the short installation path, mode selection, the three state files, profiles, and honest limitations. Detailed contracts remain linked for readers who need them.

Today, Agent Harness Kit provides an installable command-line interface, operating contracts, roles, templates, playbooks, capability-based routing, bounded execution, validation, and packaging. It is a scaffold followed by capable agents, not a background daemon. It does not promise to create chats, worktrees, merge branches, deploy, or publish on platforms that do not expose and authorize those capabilities.

The project uses the MIT License and is open to the community. In short: install the Kit, choose standard or hackathon delivery, approve the context once, and let the agent work from durable state toward a testable result.
