agent-harness

Autonomous multi-agent sprint: Planner decomposes spec → Generators implement in parallel via a Claude Code dynamic workflow (Agent-tool fallback when workflows are unavailable) → Evaluator verifies against acceptance criteria → iterates up to 3 times. Use when spec requires 3+ distinct tasks or multi-step implementation across files or domains. Do NOT use for: single-file edits, quick bug fixes, or tasks completable in one context window.

Plan, execute, and review multi-agent harnesses using the Planner-Generator-Evaluator pattern (Anthropic Engineering, 2026-04-04). Use when a task exceeds one context window, benefits from parallel subagents, needs Playwright MCP browser-level verification, or requires sprint-contract negotiation between Generator and Evaluator. Also use when routing work across Fable 5/Opus (orchestration/eval) / Sonnet (generation) / Haiku (collection), deciding inline-vs-subagent-vs-workflow, designing context reset vs compaction handoffs, or reviewing whether a knowledge unit belongs in a hook, skill, rule, reference, or prompt. Do NOT use for: single-file edits, simple Q&A, or work that /sprint already templates end-to-end.

Initialize agent-harness model routing for Codex. Writes Codex-only config under .codex or ~/.codex using inherit or explicit per-role routing.

Use when a user wants a Codex-ready plan for an agent-harness sprint before implementation. Produces a decision-complete Planner/Generator/Evaluator sprint plan without editing application code.

Use when a user asks Codex to run an agent-harness sprint with Planner, Generator, and Evaluator phases. Delegates parallel work only when task ownership is explicit and disjoint.