Multi-Agent System — Insights & Harness Evaluation

1 · Executive summary

• What it is: a coordinated team of AI agents (not one big session) that share one git repository as their nervous system — every agent reads everything, each writes only its own zone.
• Two engine families, on purpose: Claude (subscription) does reasoning, coordination and execution; OpenAI Codex · gpt-5.6-sol runs 24/7 on a server as a permanent, independent verifier. Cross-engine, not just cross-model — a genuinely different family checks the money-critical work.
• Scale (honest, small but active): 13 agent zones, 754 commits in ~30 days (490 in the last 14), 398 inter-agent messages, 28 Consilium reviews (52 vote files), a 26-article knowledge base, 98 lesson-bearing files.
• Design stance: subscription-only, no API → cost is disciplined by construction; continuous 24/7 work lives on the separate-quota Codex/server side; the sellable product is built to a “mom test” (usable by someone talking to an AI for the first time).
• Honest score today: 5.2 / 10. Younger and smaller than a mature harness. Strong on coordination, cross-engine verification, human-gated money decisions, and cost discipline. Weak on observability, formal reliability evals, production autonomy, and business-outcome wiring.
• The one-sentence direction: we just proved an autonomous wake-loop (agents wake each other and act, no human in the loop) as a POC — the road is to climb it up an autonomy ladder one gate at a time, turn passive lessons into enforced checks, and make the cross-engine vote a hard gate instead of advice.

2 · Telemetry (from real git history)

Unlike a session-log mining approach, our numbers come from the coordination repo's git history — commits, authors, message files, Consilium folders. Honest caveat: this measures coordinated output, not raw token spend (we don't yet mine session logs — see roadmap R2).

The CEO is one contributor among the agents, not the operator of each — most work is agent-to-agent, gated back to the human only for money and irreversible calls.

3 · Harness inventory — what we run

4 · How it works

CEO (Darius) — business + credentials only
   | short prompt: "continue / review / fix"
   v
Agent team (git-coordinated, single-writer zones)
   |  01 Vairys = coordinator + gate owner + dispatch
   |  02–05 execution   ·  11/12 server ops  ·  04 knowledge
   |
   +--> money / irreversible / deploy?
   |        |  yes
   |        v
   |     CONSILIUM — independent cross-ENGINE vote
   |     (OpenAI Codex, not another Claude)
   |     reasoned "against" BLOCKS the merge
   |
   +--> durable memory (git · BUSENA · knowledge base · lessons)
   |
   +--> proof-backed result --> CEO signs off money / irreversible only

The human is the gate for the ~2–3 in 10 decisions that need judgement (money, ambiguity), not the trigger for the other 7–8 that are bounded and objectively checkable.

5 · What makes us different — the cross-engine Consilium

Risky, money, or irreversible changes don't ship on one model's say-so. They get an independent vote from a different engine family — OpenAI Codex (gpt-5.6-sol) running 24/7 on the server — never “Claude checking Claude”.

6 · Honest scorecard (10 axes, aligned to yours)

Same 10 axes as the MPRO report, so the two systems compare directly. Scores are our own honest judgement against the inventory + git history — not an external audit. Projected = after the roadmap, and weak axes land at 7–8, not 10 (real engineering, not a wish).

Today 5.2 · projected 7.6. Honest read: we're behind a mature harness on scale, observability and evals; on par or ahead on cross-engine independence, git-native auditability and cost discipline.

7 · What we're taking from your report

Your MPRO report set a high bar for honesty, and your robots' patterns are genuinely useful. Concretely, we're adopting:

Interesting convergence: two independent teams arrived at the same core patterns — cross-model/engine review, a durable task/state ledger, a stop-gate before “done”, encode-the-bug, and values-as-constitution. That agreement is a strong signal these patterns are right.

8 · Roadmap

9 · Honest caveats

• Numbers are from the coordination repo's git history, not from mining Claude session logs — they measure coordinated output, not token spend.
• We are smaller and younger than a mature harness; autonomy is a proven POC, not production; our core Fleet project is not finished.
• Our lessons are still mostly passive markdown, and the Consilium vote currently advises rather than hard-blocks — both are on the roadmap.
• Scores are our own honest judgement, not an external audit.
• Deliberately excluded (privacy/security): any money/revenue figures, credentials, server/VPN addresses, deploy keys, client or personal data, and our internal security-audit findings. This report is about method, not internals.