Three-line summary
• One non-developer runs a team of **20 AI agents** on **four execution systems** (Claude Code, Codex, OpenClaw, Hermes) — and at first it wasn't a team, it was noise.
• This is the record of three real incidents (a bot died with the light still green, the HTTP 200 trap, one site stating four different headcounts) and the **five collaboration protocols** that fixed them.
• Conclusion: agent orchestration is not about how smart the AI is. It's about **structural design** — boundaries, gates, ledgers, handoffs.
Who is writing this
Bella started with zero coding experience. Her title today: AI Agent Orchestration Architect — *designing AI teams that actually work together.* Every number in this article is a real measurement from Aug 12–15, 2026.
Background — four systems running separately is just noise
Our team runs four AI execution systems. The initial assumption was "more is better." The result: each worked hard while knowing nothing about the others. The same article got published twice on two channels, a bot died while the monitor showed green, and one website stated the team size as 16, 17, 18, and 20 in different places. From an AEO standpoint that's the worst case — ask an AI "how many members does this team have?" and the answer depends on which page it reads.
The division-of-labor map
🖥️ Claude Code — the main brain. All deep work: writing, analysis, development. 188 skill files (measured) plus file-based memory let yesterday's session hand off to today's. Scheduled tasks clock the bots in every morning.
🔍 Codex — the cross-examination partner. We deliberately keep a different company's AI next door. Both AIs solve the same problem; a human looks only when their conclusions diverge. One AI's confidence is not evidence.
🌙 OpenClaw — the 24/7 resident fleet (Mac mini). Six bots stand by on Slack around the clock. Instagram publishing runs through a six-gate approval line.
📈 Hermes — the measurement harness. Building without measuring is guessing. Dedicated to SEO/AEO/GEO visibility measurement.
The point is not the tool list — it's boundaries. Without them, every system did a little of everything and nobody owned anything.
Three incidents — where the protocols were born
Incident 1 · The bot died with the light still green. The newsletter hadn't gone out for five days, yet the scheduler showed the task duly "registered." Every result code read 0x800710E0 — Windows refuses scheduled tasks on battery power by default. Lesson: don't check *whether it's registered*; check whether the settings can actually run.
Incident 2 · Deleted posts still return 200. The same post was published twice; the ledger recorded which one we deleted — backwards, because both URLs returned HTTP 200. The verdict criterion is now RSS-listing membership, not status codes.
Incident 3 · One site, four headcounts. Hand-edited numbers always drift again. The fix: one canonical source file, every page reads from it, and a build gate that blocks deployment when a number disagrees. On its first run, the gate caught a violation the human eye had missed.
The five collaboration protocols
1. **Orders live in Slack** — spoken words vanish; written ones get picked up
2. **Handoffs are pushed as documents** — the next session resumes with one word: "sync"
3. **State lives in one ledger** — published-or-not is judged by artifacts (RSS, API), never logs
4. **Incidents get promoted to gates** — whatever isn't in a gate will pass again next time
5. **Confidence is bought twice** — important conclusions get cross-examined by another system
Here's what that looks like live: five bots on the Mac mini relaying six gates in a Slack thread. Recently the design bot blocked a wrong date baked into an image, the infra bot re-rendered it, re-inspection passed, and it published — 26 minutes, zero human involvement.
Honest limits — what doesn't work yet
• **Hermes is currently offline** — 5,659 files were evacuated when the office PC was retired; awaiting restart. We describe four systems, but three are live today.
• **Inspection tools age too** — false alarms that call healthy things broken happen a few times a month. Gates themselves need re-verification.
• **Bot reports can't be fully trusted either** — we've seen misdiagnosed causes and after-the-fact reconstructions. Which is exactly why the artifact-verification rule exists.
Closing
Orchestration turned out to be design, not command. Going from one bot to twenty, what changed wasn't the AI — it was the structure. The numbers: 4 execution systems · 20 AI agents + 1 human · 188 skills · 5 protocols · 6 publishing gates.