$ ls /agents/ | wc -l → 37
/agents fleet
Sub-agents are spawned via the Agent tool with subagent_type=<name>. Every non-trivial spawn forks its own git branch, writes a 6-file artefact bundle, and registers a row in kei-ledger (RULE 0.12).
tiers
| T1 validators | 6 read-only audit gates |
| T2 researchers | 10 read-only analysis |
| T3 implementers | 13 writes code/files |
| T5 orchestrators | 8 hub-and-spoke |
# spawn from main session Agent({ subagent_type: "validator", prompt: "verify the 4 cites in...", isolation: "worktree" })
validator
RULE 0.4 enforcement gate — fact-checker and hallucination detector. Verifies API existence, version compatibility, documentation claims, code reality, and external benchmarks. Read-only — emits VERIFIED / UNVERIFIED / FALSE / PARTIALLY TRUE per claim.
validator-api
Verifies API existence and signatures. Reads docs, greps source, fetches OpenAPI / vendor reference. Read-only.
validator-benchmark
Verifies external benchmark claims (p50/p95/throughput). Read-only.
validator-code-reality
Verifies behavioural claims against running code. Reads tests, traces, recent runs. Read-only.
validator-doc
Verifies documentation claims against code reality. Read-only.
validator-version
Verifies version compatibility (semver, MSRV, transitive deps). Read-only.
git-model (RULE 0.12)
# on fork, kei-ledger writes: agent_id = "a3f9c1..." branch = "agent/code-implementer-1714291200" parent = "feat/site-build" spec_sha = "sha256:7d2a..." status = "running" # 6-file artefact bundle in .claude/agents/<id>/ spec.md · plan.md · progress.json chatlog.md · handoffs.md · review.md