Agent Fleet Architecture
Current lanes, naming, orchestration, and cleanup options.
Canonical Names
| Agent | Host | Runtime | Role |
|---|---|---|---|
max-mini-hermes | Max Mac mini | Hermes | Main agent. Telegram and primary Matrix route. |
max-mini-claw | Max Mac mini | OpenClaw | Mac worker lane, UI and local action lane. |
lera-mini-hermes | Lera Mac mini | Hermes | Backup worker and clean second lane. |
lera-mini-claw | Lera Mac mini | OpenClaw | Clean review, QA, and implementation lane. |
max-vps-hermes | Contabo VPS | Hermes | Ops-only production lane. |
max-macpro-hermes | MacBook Pro | Hermes | Local high-context controller. Gateway normally stopped. |
Current Shape
TalkMax uses Telegram or main Matrix.
Route
max-mini-hermes classifies and keeps thread light.ReserveManager leases host slots before workers start.
ExecuteHermes or Claw workers run in isolated lanes.
ReportWorkers write handoff; main agent returns synthesis.
How Agents Connect
Telegram or Matrix
-> max-mini-hermes
-> manager/scripts/agent-team-dispatch
-> agent-host-lease reservation
-> agent-team-launch
-> worker session with --source tool and --worktree
-> handoff file, Matrix checkpoint, wiki artifact
-> concise answer back through max-mini-hermes
Room Policy
- Ops room: free-response for that one agent.
- Project room: mention-only.
- Contabo/VPS ops:
max-vps-hermesfirst responder. - Together mode: ask main Hermes to run a team.
- Mac actions: approval before broad local changes.
Live State
max-mini-hermes
gpt-5.5 / OpenAI Codex
Gateway running. Telegram configured. Main lane.
max-mini-claw
runtime still says max-mini-openclaw
Gateway running. Canonical Matrix slug active.
lera-mini-hermes
gpt-5.5 / OpenAI Codex
Gateway running. Matrix lane.
lera-mini-claw
runtime still says uniclaw-openclaw
Gateway running. Canonical Matrix slug active.
max-vps-hermes
MiniMax in Matrix
VPS Codex CLI works; Matrix Codex path needs fix.
max-macpro-hermes
gpt-5.5 / OpenAI Codex
Local controller. Gateway installed but stopped.
Architecture Options
| Option | Shape | Score | Use When | Risk |
|---|---|---|---|---|
| A. Main router plus split lanes | Talk mostly to max-mini-hermes. It dispatches bounded workers. |
9/10 | Best default now. | Claw slug migration complete. Worker handoff discipline enforced by agent-team-launch. |
| B. One Matrix command room | Every agent listens in one shared room. | 5/10 | Quick demos only. | Noisy, weak permission boundaries, easy duplicate work. |
| C. Telegram-only front door | Max talks only to main Hermes; everything else is hidden. | 6/10 | Simple daily use. | Less observable; harder to debug agent-specific failures. |
| D. Manager hard router | Manager owns queues, leases, tickets, dispatch, and policy. | 8/10 later | After health checks and restart gates are mature. | More code and more state to keep correct. |
Final Migration
- Create or alias Matrix identities for
@max-mini-clawand@lera-mini-claw. - Update OpenClaw configs, allowlists, room maps, and service env.
- Rename Ops rooms to
Ops - max-mini ClawandOps - lera-mini Claw. - Remove stale
maxiclaw-*,uniclaw-*, and*-openclawentries after smoke tests pass. - Keep SSH aliases
maxiclawanduniclawworking until scripts are migrated.
Main Thread Fix
- Main Hermes should not execute long work inside Max's chat thread.
- VPS-local ops should start with
max-vps-hermes, then escalate. - Worker commands should use
hermes chat -Q --source tool --worktree. - Every worker needs a lease, allowed path, success criteria, and handoff path.
- Main thread receives only accepted, blocked, checkpoint, and done summaries.
- Wiki pages hold durable artifacts; Matrix holds operational messages.
Source Links
- Max Wiki
- Social Poster wiki area
- Synapse user registration API
- Matrix Client-Server API
- Canonical private source:
manager/brain/shared/ops/agent-fleet-ops.md