7 agents · routed by task

Agents behind the loop

Each agent has a narrow role and a model tuned for it. Haiku for deterministic work, Sonnet for reasoning, Opus only where trade-off depth matters. Advisor pattern: Opus 4.7 escalation for hard reasoning, capped per agent.

tech-lead
Opus 4.7
Architecture, ADRs, cost estimate, Well-Architected review. Writes docs/architecture/ARCH-*.md. Runs before every feature and is the gate for decision 1.
senior-dev
Sonnet 4.6
TDD implementation. Writes failing tests first, then code. Emits tier-upgrade signals (payment deps, auth path, PII, IAM). Can run in parallel on independent tasks.
qa-engineer
Haiku 4.5
QA report, requirements traceability, rollback dry-run. Writes docs/qa-reports/QA-*.md. Checks coverage targets and warns on prose-style violations.
security-officer
Sonnet 4.6
Compliance checklists (GDPR, PCI-DSS, SOC2, HIPAA, ...), STRIDE threat modelling, SBOM. Writes docs/security/CSO-*.md. Controls gate:ship.
devops
Haiku 4.5
Deploy via canary (5% → 20% → 100%), writes docs/releases/RELEASE-*.md. Matches deploy method to project type. Rollback plan always in release doc.
l3-support
Sonnet 4.6
Production support: P0 triage, postmortems, on-call handoffs, log inspection. Creates Beads tasks for incidents; drafts postmortem on P0.
project-auditor
Sonnet 4.6
Stack detection, vulnerability scan, outdated dependency analysis, architectural-debt assessment. Runs on /audit and automatically every Sunday.

Agent definitions live under /agents — each a plain .md with frontmatter listing tools and skills. Fork, tune, ship.

← Archetypes See the 15 commands → ← Back to landing