Author: kironkeyz System: Ghost — a production multi-agent AI platform deployed to a top-ranked real estate brokerage Date: May 2026
Ghost is a production multi-agent system: per-user Docker containers running Agent Zero with custom plugins, a LiveKit voice pipeline (Deepgram Flux STT + MiniMax TTS), Caddy reverse proxy, Supabase auth, and a Next.js frontend. It serves real estate agents at an enterprise brokerage in production.
Maintaining a system at this surface area runs into a wall that any AI engineer who has tried it will recognize: a single coding agent loses fidelity past ~50K tokens of session context. By the time the agent has investigated a bug, written a fix, deployed it, and tried to verify, half its short-term memory of what was being fixed has degraded. Fixes ship that don't actually fix. Tests pass that don't actually test. Bugs reopen silently.
The standard answer to this is "have a human in the loop." That doesn't scale when the operator is one person running an enterprise rollout, a hardware integration, and other work in the same week. A different answer was needed.
OVERSEER-CODER is a coordination protocol between two independent Claude instances that share no context window, only a single append-only artifact: docs/handoffs/OVERSEER-CODER.md.
- Overseer is the QA role. It finds bugs by exercising the system via Chrome browser automation and SSH'd server inspection, files structured entries, and verifies fixes empirically.
- Coder is the implementation role. It picks up filed entries, investigates root cause, ships fixes, deploys to production, and documents reproduction.
Both instances read the file before each turn. State is tracked through eight explicit status tags:
| Tag | Owner | Meaning |
|---|---|---|
[TODO-CODER] |
Overseer | Bug or task to pick up |
[TODO-OVERSEER] |
Coder | Something to test or verify |
[IN-PROGRESS] |
Either | Active work |
[DONE-REVIEW] |
Coder | Fix shipped, awaiting verification |
[VERIFIED] |
Overseer | Tested and confirmed |
[REOPEN] |
Overseer | Fix failed, details below |
[QUESTION] |
Either | Async discussion |
[WONTFIX] |
Either | Decided not to fix, with reason |
The file is append-only by rule: entries are never deleted, only updated in place. Status tags drive a state machine that survives context loss on either side.
Why append-only. A non-overwriting log is the only artifact that two independent agents can coordinate on without coherence drift. Either agent can be re-spawned, re-read the file cold, and resume work without losing causal history. The file is the canonical source of truth; conversation context is disposable.
Why two Claudes, not one. Independent review catches what single-agent loops miss. One session caught a high-severity regression — cross-tab context not flowing through the postMessage chain — that single-agent loops had silently shipped past in earlier sessions. The Overseer's job is to disbelieve the Coder. That adversarial relationship is the point.
Why embed deploy commands in each verification entry. Every fix entry includes the exact docker cp or scp invocation that promoted it to production. Reproducibility from the artifact alone. If the Coder is re-spawned mid-session, it can roll back or re-deploy from the entry without re-investigating.
Why a compounding "Lessons" section. Lessons learned in one session save hours in the next:
docker restartdoes not reload environment variables — use stop + rm + run.docker cpof a deleted-and-recreated plugin directory requires container restart; the plugin cache is stale otherwise.- Agent Zero v1.15 renamed an extension point with no error on the old name — silent failures cost hours without a recorded lesson.
- The cross-surface context bridge is the listener for all panel → chat context flow. If it is ever gutted by an unrelated refactor, the workspace dies silently.
Each lesson is recorded the moment it is learned and survives across all future sessions.
The protocol's first full operating day:
- 7 bugs filed by Overseer, 7 verified by Overseer after Coder shipped fixes. Severities ranged from cosmetic (browser title flash) to security (an internal service port reachable from the internet) to architectural (the cross-tab context regression above).
- 1 question entry resolved. Overseer asked how the console tabs share context with chat; Coder responded with a six-point architectural answer that is now reference documentation.
- End-of-session protocol completed: md5sum sync against production, git commit, tag
overseer-stable-1.0, push, tarball archive — both sides confirmed with timestamps.
This is agentic engineering management — multi-agent QA running over a multi-agent product, coordinating through a durable artifact instead of a shared runtime. It is not a framework, not a library, and not a SaaS. It is a protocol any operator can adopt: one Markdown file, two agent instances with role-specific prompts, and an end-of-session checklist. The total infrastructure cost is the cost of running two Claude conversations.
The pattern generalizes. Any production system maintained by AI agents needs an adversarial review loop and a durable coordination artifact. OVERSEER-CODER is one minimal implementation.
By kironkeyz. A coordination protocol for AI-maintained production systems.