SDLC Harness does not depend on a specific agent. Any agent that can read repository files, run commands, and modify code can follow the same protocol.
New repositories
init generates AGENTS.md in the repository root. At the start of a task, have the agent read:
AGENTS.md and the workflow pages it references.- The current active, ready, and blocked features in
feature_list.json.
- The current feature’s
source_refs, dependencies, and verification.
- The latest session checkpoint in
progress.md.
A practical startup prompt is:
The repository already has AGENTS.md
adopt does not overwrite the file. It writes AGENTS.sdlc-harness.md instead. Add an explicit reference to the original AGENTS.md, for example:
The [agents-onboarding] check in validate verifies this discovery chain. Repository validation fails when the sidecar exists but the root protocol does not reference it.
Claude Code
In addition to the shared AGENTS.md, the template generates lightweight entry points under .claude/skills/ to help Claude Code discover the requirements, architecture, user story, feature decomposition, milestone, TDD, self-acceptance, deployment, and feedback workflows. Repository files remain the final source of truth.
Multi-agent collaboration
- Use
owner to identify the responsible party, and actor to distinguish separate agent sessions started by that owner.
claim uniqueness is always determined per feature; actor does not allow two agents to own the same feature at once.- Each agent works only on features for which it holds a valid
claim.
- Use
workspace create to isolate parallel changes in separate worktree directories.
- A review agent cannot be the
owner of the feature’s claim.
- Run
session close at the end of every session, and record the next step and any blockers in progress.md.
Before an agent reports completion, ask it to show the actual output from feature complete, validate, and session close, not just a natural-language summary.
