Agent Harness Kit, version zero point seven.

Agent Harness Kit is a local execution-governance layer for coding agents. Codex, Claude Code, or another capable host still writes the software. The Kit defines how work is understood, scoped, ordered, verified, resumed, and, when the host supports it, distributed across independent agents.

Install the command-line tool with U V, pipx, or pip, open the project, and run agent harness install. A successful installation creates the contained agent harness kit directory and two entrypoints at the project root: A G E N T S dot markdown for Codex and CLAUDE dot markdown for Claude Code. If those files already contain project instructions, the installer preserves them outside one managed bridge block. Open a new agent context at the project root and the host can discover the Kit without a separate activation prompt. Agent harness doctor verifies all three entrypoints.

Version zero point seven is adaptive. Every mutating request receives one execution lane and one independent assurance level. Direct trivial is for mechanical content or presentation edits. Vibe handles one small local behavior change with a focused deterministic check and no Harness artifacts. Graph only adds durable ordering, ownership, and resumability. Full Harness is reserved for real multi-agent coordination, a human decision loop, a required audit, insufficient model capability, unresolved consequential ambiguity, or an explicit request. Assurance none closes on executor verification. Light and full require an actual independent reviewer in a fresh context.

Before graph or full decomposition, executable preflight verifies every declared file, package script, environment-variable name, native command, validator, browser or sandbox requirement, and proven worker capacity. Missing setup becomes an exact prerequisite or capability-degradation result before task generation, instead of being rediscovered through repeated failed plans.

Resume is also evidence first. The agent begins with a bounded, side-effect-free probe of the current working tree, relevant runtime state, and the most useful available check. It reads project context, pending work, the task graph, and current handoffs only when that evidence leaves a gap or ownership must be recovered. Current tests and runtime evidence supersede stale transfer documents.

The Kit separates human state from technical state. PENDING dot markdown owns human decisions, required actions, and the macro view of unfinished product outcomes. TASK GRAPH dot markdown owns dependencies, readiness, leases, dispatch, progress, and technical recovery. This durable state allows a fresh conversation to continue without reconstructing the project from chat history.

Planning is proportional. Bounded same-context work may use a compact inline specification. Planned units normally target fifteen to thirty minutes of active agent work, with smaller or larger exceptions justified by atomicity, runtime cost, or risk. A separate TASK document exists only when another context will consume it. Handoffs and review packets are created only for a real separate reviewer, subagent, or human recipient.

Behavior changes use meaningful RED and GREEN evidence. Verification starts with the narrowest focused check, then climbs only when needed through workspace, integration, global checkpoint, and delivery. A small edit does not rerun an entire repository suite by default. Technical failures receive bounded recovery inside the approved scope; product, scope, material-cost, permission, and experimental-integrity changes still require a human decision.

When several graph nodes are ready, the scheduler considers dependencies, assurance checkpoints, worker capacity, and overlapping write or impact sets. It selects a maximum-size collision-free batch rather than a convenient greedy list. The active orchestrator uses the host's real subagent or task capability to launch that batch and refill capacity as work completes. If the host cannot provide parallel contexts, the Kit records sequential fallback instead of pretending that parallelism occurred.

Runtime state is inspectable. Graph transitions use compare and swap, temporary-file replacement, and cross-process coordination so a stale writer cannot silently overwrite a newer revision. Events form an append-only hash chain with idempotent identifiers. Run metrics record lane, assurance, execution shape, artifact count, ceremony and implementation time, approvals, effective gate hits, planned versus actual duration, reviews, remediations, graph revisions, and host-reported tokens. Repeated runs in which no gate provides value produce a visible suggestion to use a lighter lane.

The command-line interface performs deterministic installation, inspection, request routing, preflight, scheduling, state transitions, metrics, validation, and dispatch-packet preparation. The agent host performs the actual coding, subagent creation, independent review, integration, and delivery according to its capabilities and permissions. The Kit has no unattended background daemon. Leases are validated ownership contracts, not operating-system file locks. It does not grant credentials, merge, deploy, or publish on its own.

In short, Agent Harness Kit turns coding by conversation into execution with durable memory, bounded authority, proportional verification, safe coordination, and evidence of whether the process actually helped.
