Agents

harness-kit ships three agents. Each lives under .claude/agents/<name>/ with its own README, sensors, evals, guides, and skills, and is registered in AGENTS.md. All three are also invokable as sub-agents via the Task tool.

product-manager

Turns a problem into an engineering-ready spec. Two artifacts, two skills:

prd: Product Requirements Document, business-facing.
prp: Product Requirements Prompt, the engineering handoff.

Entry: /product-manager:run (full PRD → PRP), or /product-manager:prd / :prp solo. Gated by prd-structure/prd-quality and prp-* sensors and evals. Can publish to Confluence when JIRA_USERNAME + JIRA_API_TOKEN are set.

staff-software-engineer

Turns an approved PRP into a merged PR. Picks the right area skill from the repo's files:

backend, web, mobile, devops: discipline conventions, overridable per repo via .claude/conventions/{area}.md.
designer: a cross-cutting skill applied when building a new UI: Material Design 3, dark/light theme, modern type, Behance-grade polish, i18n (en, pt-BR, es), context-aware favicon. Layers on top of the area skill. See Designer Skill.

Entry: /sse:run (plan → dev → test → pr → monitor), /sse:run --local (no PR), or the single-stage commands. The SDD variant /sse:sdd runs a spec-driven loop (plan once, dev↔test↔eval until the PRP is satisfied, cap 3 iterations, no auto PR). See Pipeline and Stages.

system-architect

Turns a system/problem into a rigorous System Design Doc, then runs an adversarial design review. Like the SSE agent picks an area skill, this agent picks a topic skill: one per classic system design problem, drawn from the System Design podcast series:

design: generic fallback for any system.
review: adversarial staff-level review of an existing design.
url-shortener (#1), rate-limiter (#2), search-engine (#3): topic playbooks.

Entry: /system-design:run (design → review), /system-design:design, /system-design:review. Gated by design-structure/design-rigor sensors and design-quality/design-review-depth evals. The deep theory and references for each topic live in this wiki. Method: System Design Method.

All three follow the same harness shape, guides (feedforward), sensors + evals (feedback), an approval marker per artifact, token accounting per phase, and a skill the orchestrator dispatches to. The product-manager and staff-software-engineer agents chain into the six-stage golden path; the system-architect agent is an optional front stage before it.

v5: from monoliths to orchestrator + leaves

Today each agent runs its stages inline, and the sensors and evals run in the same context as the author. v5 decomposes this into an orchestrator (the main session, owning state and gates) that dispatches small, single-purpose leaf subagents: an intake harvester, per-stage authors, adversarial evaluators that never wrote the artifact they grade, and parallel reviewer panels. The agents above remain the conceptual owners of their stages; internally, each stage becomes a set of leaves the orchestrator coordinates. See Orchestration and Subagents and Autonomy.

Routing

When you type a slash command the entry point is unambiguous. When you describe work in natural language, the main session consults the routing table in AGENTS.md:

Intent	Route
idea → merged PR	`/golden-path`
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 with bundled assets).
Register it in AGENTS.md under the right section.
Add .claude/commands/<name>.md if it is slash-invocable.
Wire lifecycle hooks under .claude/runtime/hooks/<name>/ if it needs them.