Agents
harness-kit ships three agents, registered in AGENTS.md. Each lives under .claude/agents/<name>/ with its own README, guides, sensors, evals, and skills, and each is also callable as a subagent through the Task tool.
product-manager
Writes the PRD (business-facing) and the PRP (the engineering handoff). Run /product-manager:run for both, or /product-manager:prd and /product-manager:prp alone. Sensors prd-structure, prd-acceptance-criteria, prp-structure, prp-context-quality, and prp-links gate structure; evals prd-quality, prd-readiness, prp-quality, and prp-context-readiness gate quality. With JIRA_USERNAME and JIRA_API_TOKEN set, it can publish to Confluence.
staff-software-engineer
Takes an approved PRP to a merged PR. It picks an area skill from the repo's files: backend, web, mobile, or devops, each overridable per repo with .claude/conventions/{area}.md. The designer skill layers on top when you build a new UI; see Designer Skill.
Run /sse:run for plan, dev, test, pr, and merge monitor, /sse:run --local to stop before the PR, or any single stage. /sse:sdd plans once, then loops dev, test, and a spec-satisfied eval until the PRP is met, capped at 3 iterations, with no automatic PR. See Pipeline and Stages.
system-architect
Writes a System Design Doc, then runs an adversarial staff-level review that returns ship, revise, or block. It picks a topic skill per classic problem: url-shortener (#1), rate-limiter (#2), search-engine (#3), with design as the generic fallback and review for the review pass.
Run /system-design:run, /system-design:design, or /system-design:review. Sensors design-structure, design-rigor, and review-structure and evals design-quality and design-review-depth gate it. The method is on System Design Method.
The shared shape
All three use guides as feedforward, sensors and evals as feedback, an approval marker per artifact, and token accounting per phase. product-manager and staff-software-engineer chain into the six-stage golden path; system-architect is an optional front stage.
Each stage runs as an orchestrator plus leaf subagents. The main session owns state and gates and dispatches an intake harvester, per-stage authors, and evaluators that never wrote the artifact they grade. The evaluator is a fresh Claude subagent or Jev, per /hk:eval; see Evals and Orchestration and Subagents.
Routing
A slash command picks its entry point. For plain-language requests the main session reads the routing table in AGENTS.md:
| Intent | Route |
|---|---|
| idea to merged PR, hands-off | /pipeline:run "<idea>" |
| idea to merged PR | /golden-path |
| harvest context | /intake:run |
| draft a spec | product-manager |
| ship an approved PRP | staff-software-engineer |
| design a system at scale | system-architect |
Adding an agent
Create .claude/agents/<name>.md, or <name>/agent.md when it bundles assets, and register it in AGENTS.md. Add .claude/commands/<name>.md if it takes a slash command. Put lifecycle hooks under .claude/runtime/hooks/<name>/ and wire them in .claude/settings.json.
See also Pipeline and Stages, Golden Path, Harness Engineering.
This page mirrors Agents in the wiki. Edit it there.