factory/vibe/PRD/ai-back-office/agent-catalog.md

vibe > PRD > AI back-office > Agent catalog

Agent catalog — who does what, with which brain, under which conductor

Status: In design (assignments are hypotheses until proven) Last Updated: 2026-07-11 Up: AI back-office hub Related: Task inventory · Model fleet · Agent architecture · QA strategy

An "agent" here is the concrete triple prompt + model + orchestrator bound to a task. This page names the classes, assigns every task, states how each assignment gets proven (éprouvé), and fixes the syntax of the document surface agents read to do the work. Honesty first: several agents are deliberately LLM-free — a cron-driven script with validators is the best "agent" for deterministic work, and the prompt column says so.

Agent classes

Seven prompt skeletons; every atom's prompt extends exactly one. Skeletons live with the fleet code (fleet/classes/<class>.md — see D8).

Class	Prompt skeleton (the invariant part)	Model policy	Orchestrator	Serves
Sentinel	closed-set classification, schema-constrained output, refuse below threshold	Pi SLM (GBNF) → M4/Mistral fallback	k3s CronJob → queue	T01, deadline detection
Extractor	document → JSON Schema, zero tools, dual independent run, never "fix" arithmetic	M4 local ∥ Mistral (agreement), Claude escalation	queue workers (cluster leg + hermes leg)	T02, T16
ERP scribe	manifest assembly over the write skills, ledger grammar honored, predicted-delta before card	Claude (Agent SDK headless)	gateway handler → gate → promote	T03, T04-create, T05, T08-ambiguous
Deterministic controller	— (no prompt: scripts + validators + linter)	—	k3s CronJobs	T07, T08-matched, T11, T13, T14
Analyst-writer	narrative strictly over verified figures; cite from ERP/journals only; no advice	Claude, or M4 for local prose	crons → digest	T06 drafts, T09, T10 narrative, T15 exceptions
Researcher	sourced-claims-only (official domains), effective dates mandatory, output = diff proposal	Claude + web	quarterly / event-driven	T12
Knowledge archivist	vault conventions: append-only, idempotent frontmatter, PARA filing hints	per vault doctrine (Ornith/Mistral/Claude)	hermes cron + per-run hooks	T17

Assignment matrix

Status legend: ✅ proven in operation · 🧪 built or designed, to prove (à éprouver) · ⬜ not built. Where a task splits (deterministic core + LLM edge), both appear.

Task	Class	Prompt / code	Model	Orchestrator	Proof gate	Status
T01	Sentinel	fleet/atoms/mail-classify/	Pi Qwen3-class 1.7–4B, GBNF	k3s CronJob (30 min)	POC-2 ≥ 95 % on 200 labeled mails	🧪
T02	Extractor ×2	fleet/atoms/invoice-extract/	M4 structured ∥ Mistral JSON; Claude escalation	queue workers	POC-1+POC-5 ≥ 98 % critical fields	🧪
T03	ERP scribe	dolibarr-sandbox-write + fleet/atoms/invoice-record/	Claude headless	gateway → gate → promote	POC-1 10-invoice exit gate	🧪 (write skills ✅, loop ⬜)
T04	Controller + scribe	dolibarr-thirdparty-completeness (audit, no LLM); creation rides T03	— / Claude	monthly CronJob / with T03	audit: live now; creation: POC-1	✅ audit · 🧪 creation
T05	ERP scribe + auditor	dolibarr-recurring-templates + dolibarr-invoice-audit + fire atom	Claude	monthly cron (1st) + card	first agent-fired invoice == manual twin (D3)	⬜
T06	Analyst-writer	fleet/atoms/dunning-draft/ (+ T17 retrieval for tone/history)	Claude	weekly cron → card	N consecutive drafts approved unedited	⬜
T07	Controller	arcodange-bank-reco (bank-match.sh)	— no LLM	weekly CronJob	fixture-proven; standing zero-delta invariant	✅ skill · 🧪 standing
T08	Controller + scribe	payment-record.sh manifests from matched movements	— matched; Claude ambiguous	queue → gate → promote	POC-3 one clean month	🧪
T09	Analyst-writer	balances workflow + fleet/atoms/cash-report/	figures deterministic; M4 prose	monthly CronJob + hermes leg	figures == live bank APIs, every run	🧪
T10	Controller + analyst	dolibarr-tva-summary + narrative atom	— figures; Claude narrative	calendar-triggered (T11)	POC-4 €-parity vs filed	🧪 (skills ✅)
T11	Controller	calendar file + fleet/atoms/deadline-remind/	— no LLM (parsing upstream in T01/T12)	daily k3s cron → gateway	synthetic-calendar firing test	⬜
T12	Researcher	fleet/atoms/reg-watch/	Claude + web	quarterly + event	every claim sourced + PR review	🧪 (method proven authoring this PRD)
T13	Controller	dolibarr-data-snapshot	— no LLM	daily CronJob + around writes	drift alert fires on seeded change	✅ skill · 🧪 cron+alert
T14	Controller	ops/backup + freshness watchdog	— no LLM	daily CronJob (live) + monthly drill	restore drill green monthly	✅ backup/restore · 🧪 watchdog+drill cadence
T15	Composer + analyst	skill composition + exception narrative	— checks; Claude narrative	monthly CronJob → digest	first pack matches a manual cohort review	⬜
T16	Extractor	fleet/atoms/doc-file/	M4 local	per-document queue	document golden set	⬜
T17	Knowledge archivist	sb.py + hermes second-brain skill + capture atom	vault doctrine (Ornith/Mistral/Claude)	hermes cron (live) + per-run hooks	deposits idempotent over re-runs; retrieval dated	✅ vault side · 🧪 fleet side

Proving protocol — how a 🧪 becomes a ✅

The matrix is a set of falsifiable hypotheses, not documentation:

1. Tier choice is proven by POC-5's bench (accuracy × latency × cost on the golden set) and recorded into the atom's model_policy — if the Pi can't hold T01's bar, the matrix cell changes, not the bar.
2. Loop viability is proven by the owning POC's exit criterion; autonomy then follows the promotion gates.
3. In operation, the matrix is re-scored from run journals: every escalation and tier fallback is journaled, so the monthly ops review reads which tier actually served each task. A cell whose fallback rate exceeds ~20 % is wrong and gets reassigned.
4. A ✅ is revocable: incident → demotion → the cell reverts to 🧪 with the same path back.

The document surface agents read

What an agent knows about this system, it learns from files. Their syntax is part of the architecture:

File	Read by	Lives at	Syntax rules
AGENTS.md	every agent, session start	each repo root	Orientation map: what the repo is, operating rules, pointers — the factory AGENTS.md is the canon (diagram + tables + hard rules). The erp repo's must gain a Fleet section: environment rules, ledger-grammar pointer, registry location. Keep it short; link, don't inline.
SKILL.md	Claude Code (auto-discovery), hermes (snapshot)	.claude/skills/<name>/, ~/.hermes/skills/<domain>/<name>/	YAML frontmatter name + description; the description carries the triggers: capability summary + explicit "Use when…" and "SKIP for…" clauses (the proven dolibarr-* pattern). Body = numbered workflows; executables under scripts/; secrets in mode-600 gitignored .env.
atom.yaml (registry)	orchestrators, humans, CI	fleet/atoms/<atom>/	The contract fields: I/O JSON Schemas, invariants, side_effect_class, idempotency key, earned autonomy + eval evidence link, model_policy. Folder name = atom name = registry name — the <app> join-key discipline applied to atoms.
prompt.md	the model, at runtime	next to atom.yaml	≤ ~40 lines: role (1 line), task, output = reference to the schema (never a prose re-description), refusal/escalation clause. No business rules in prompts — rules live in the fiscal profile and validators (code); prompts stay thin, versioned, diff-reviewable. Extends one class skeleton (fleet/classes/).
fiscal.yaml + calendar.yaml	fiscal atoms, T11	fleet/profile/	Schema'd YAML; every rule carries an effective_from (and effective_until when known); mutations arrive as PRs (T12 proposes, human merges).
Runbooks	humans + agents	factory vibe/runbooks/	House rule: every step marked [AGENT] (safe, delegable) or [HUMAN] (prod-mutating, approval-bound) — the same markers as the task inventory.
.mcp.json	agents needing MCP tools (vault, ERP)	repo/vault roots	Servers declared with env-var indirection for keys (${OBSIDIAN_API_KEY} pattern) — never literals.

Cross-cutting rules: English for all agent-facing files (house language policy); write descriptions for retrieval — agents discover skills by their description text, so triggers belong there, not in the body; same-change freshness — a change to an atom that leaves its SKILL.md/atom.yaml/prompt.md stale is an incomplete change (the guidebook-maintenance rule extended to the fleet); one capability per file; frontmatter over prose for anything a machine parses.