Skip to content

AGENT-ROSTER — Manifest of Agent Types and Waves

Status: live · Version: 1.0.0 · Camada: 12 (Ecossistema de Agentes Especializados) Data source: docs/reference/registries/_agents-registry.yaml (current: 25 agents) Target: 35 agents in 4 implementation waves per GUIA v3.2.2

Current roster (25 agents — v1.0)

Layer-4 specialists (20)

Agent id Domain Primary orchestrator Status
frontend-specialist client-side UI frontend_orch active
mobile-specialist mobile apps frontend_orch active
backend-node-specialist server (Node) backend_orch active
backend-python-specialist server (Python) backend_orch active
data-engineer-specialist DBs, vectors, search backend_orch active
ai-agent-specialist LLM agent assembly ai_ml_orch active
vision-specialist image/CV ai_ml_orch active
audio-specialist speech/music/voice ai_ml_orch active
video-specialist video gen/edit ai_ml_orch active
devops-specialist CI/CD/IaC devops_orch active
observability-specialist obs stack devops_orch active
qa-specialist test creation qa_orch active
security-specialist security review security_orch active
auditor-haiku universal cheap auditor any always-on
quant-analyst-specialist finance models finance_trading_orch active
trader-specialist exchange integration finance_trading_orch active
payments-specialist Stripe/PayPal integrations_orch active
communications-specialist Slack/Discord/WhatsApp integrations_orch active
planner-opus deep planning cortex use sparingly
doc-writer-specialist technical writing meta_orch active

Layer-5 workers (5)

Agent id Capability Notes
code-writer-worker write_file / edit_file atomic
file-operator-worker read / list / glob / grep atomic
api-caller-worker HTTP with retry/backoff atomic
test-runner-worker run_command / parse output atomic
git-worker git operations restricted to status/diff/log/commit

Target roster (35 agents — full v2.0)

Wave 1 — Foundation (current; shipped in v1.0)

✅ All 20 specialists + 5 workers above (25 agents)

Wave 2 — Architecture & Operations (planned, +4)

  • architect — system architect (proposed L4 specialist for BLUEPRINT/ADR work)
  • release-engineer — go-live coordinator (currently humans + devops)
  • on-call-responder — first-page handler (currently humans)
  • incident-commander — IC for SEV1/2 (currently humans)

Wave 3 — Knowledge & Learning (planned, +3)

  • memory-curator — promotes episodic → semantic, prunes stale
  • knowledge-decay-detector — finds outdated semantic memories
  • pattern-miner — runs PATTERN-MINING (Camada 11)

Wave 4 — Cognitive (planned, +3)

  • mirror-learner — implements MIRROR-LEARNING-SPEC (Camada 18)
  • dream-replay-agent — sleep-cycle consolidation
  • strategy-feedback-agent — vagus nerve (execution → strategy)

Total at v2.0: 25 (current) + 4 + 3 + 3 = 35 agents

Gap from current (25) to target (35)

10 agents pending implementation. None are blocking v1.0; all are post-runtime work.

Specialization rules

  • Each specialist owns a CLEAR domain (no overlap)
  • Each specialist binds 5–10 preferred_skills (focused subset)
  • Specialists may not invoke other specialists directly — must go via orchestrator (Rule 6)
  • Workers are atomic; specialists never bypass them for the worker's capability

Sizing rationale

Why 35 and not 50 or 20? - 20 is the floor: cover all top-level technical domains - 35 includes meta-roles (release, memory curation, cognitive) - > 35 risks over-specialization (one-trick agents that mostly idle) - The orchestrators (12 total) keep the agent count from exploding combinatorially

Adding a new agent

  1. RFC describing role + non-overlap with existing
  2. ADR formalizing the decision (Rule 7)
  3. Entry in _agents-registry.yaml (lean index)
  4. Manifest in _framework/agents/<id>/MANIFEST.md (instance, when that subdir is populated)
  5. ONBOARDING (28-step) executed before the agent accepts capsules
  6. Update this roster + bump version

Retiring an agent

  1. Deprecate notice: 30 days minimum
  2. replaced_by: <id> set in registry
  3. POSTMORTEM if retirement was due to incident
  4. Move row to "Retired" section of this doc

Cross-references

  • docs/reference/registries/_agents-registry.yaml — lean index
  • docs/reference/templates/AGENT-MANIFEST.md — per-agent format
  • docs/reference/specs/AGENT-SPEC.md — what an agent IS
  • docs/reference/specs/ORCHESTRATION-SPEC.md — hierarchy
  • WAVES (per release) — sequencing of new agents into roster