Agent Harness Kit.

Agent Harness Kit is an installable scaffold for organizing software development with coding agents. It turns context, decisions, tasks, verification, and handoffs into durable artifacts, so work can continue without depending on chat memory, rescanning the whole repository for every request, or falling into endless implementation and review loops.

Install it with U V, pipx, or pip, then run agent harness install inside the project. The installer creates a contained Kit directory and safe root entrypoints. Codex enters through A G E N T S dot markdown; Claude Code enters through CLAUDE dot markdown. Both follow the same context, graph, and platform-neutral contracts.

On first use, the agent briefly presents the Kit and begins discovery before proposing technology, architecture, or design. In an existing project, it reuses evidence already available. A new feature automatically triggers focused discovery to clarify users, access, authentication, empty states, failures, recovery, and success criteria. Technical planning starts only after that direction is approved.

Truly simple requests do not pay the full workflow cost. Changing one button color, correcting copy, or adjusting spacing can go directly to the file with the smallest useful check. When work introduces behavior, business rules, data, dependencies, risk, or broader impact, the Kit creates an executable SPEC. Non-trivial work receives one concise writing plan and small tasks, normally targeting two to five minutes, with explicit outcomes, allowed files, acceptance criteria, and verification. The agent executes the SPEC instead of improvising missing behavior.

Code behavior changes and bug fixes use test-driven development. First, a meaningful test fails for the intended reason. Next, the minimum production code reaches green. Finally, the agent refactors inside the approved scope and runs proportional regression. RED and GREEN evidence stays in the handoff.

At the start of a context window, on resume, or for project status, the reading order is mandatory: approved context, PENDING dot M D, then TASK GRAPH dot M D. PENDING owns human decisions, human actions, and the macro view of what remains unfinished. TASK GRAPH owns technical order, dependencies, leases, contexts, and execution. Status reports stage, progress, automatic work, human and technical pending items, blockers, next action, and inspectable paths.

Model routing now happens at real dispatch time. Mechanical work may use an economical model, normal implementation uses a balanced model, and consequential architecture, security, or difficult integration can use a frontier model. The adapter resolves a model and reasoning effort actually exposed by the host, applies them when creating the task or subagent, and records the platform response. A model name written in a file without dispatch confirmation is not executed routing.

Codex also has native agent dispatch. The Kit selects the neutral role, constructs only the required context packet, creates a fresh subagent without inheriting the conversation, and records identity, context, model, reasoning, and host response. Implementer and reviewer remain separate. When subagents are unavailable, implementation uses an explicit sequential fallback; review never pretends to be independent in the same context.

When two or more independent tasks are ready and the host proves parallel capacity, the scheduler selects a safe batch. It respects dependencies, assurance, and file collisions, launches subagents without waiting for one another, and immediately refills the first released slot. Parallel branches converge only through an explicit integration task. If the host cannot provide real parallelism, the Kit records sequential execution instead of simulating concurrency.

When declared checks pass, the task is completed and reported without ceremonial approval. The next ready task may start. An independent reviewer launches automatically in a fresh context and judges implementation against the SPEC, not the prompt or implementer memory. There is one initial review and, only for a real blocker, at most one focused re-review of the correction and related regressions. There is never a third round.

Implementation attempts, no-progress cycles, and context expansions are also bounded across one goal lineage. Changing the model, agent, task, or session does not reset those counters. At a ceiling, the agent records evidence and replans instead of continuing to consume tokens without progress.

The Kit supports standard delivery and hackathon mode. Hackathon mode compresses discovery, prioritizes one complete demonstrable slice, can split independent work by area, and aims quickly for a testable M V P. Learning can be added to either pace, but only after the user approves the exact note destination, such as Markdown, Obsidian, or Notion with the required connector.

In repositories that already contain rules, agents, or another harness, nothing is overwritten silently. The Kit uses inventory, provenance, namespaced coexistence, and human review before replacement. Leases are validated contracts, not operating-system locks. The Kit does not grant credentials, merge, deploy, or publish without separate capability and authorization.

In short: install Agent Harness Kit, choose the pace, approve the context, and let specialized agents work through SPECs, tests, safe parallelism, independent review, and durable state toward a verifiable result.
