Practices / Agency

Agency AI Coordination Playbook

Coordination practices for agencies running AI agent teams that manage client advertising, call centers, project delivery, and sales pipelines. Battle-tested patterns from a 25-agent production deployment.

4 practices 10 categories
Download CLAUDE.md Generate Custom CLAUDE.md
All (4) Escalation (4) Operations (4) Client Management (4) Communication (4) Data (4) Sales (3) Learning (3) Architecture (3) Ownership (3) Call Center (2)

Communication

Rule

Agent Message Bus

Direct agent-to-agent communication via structured inbox files. Each agent has an inbox at a known path. Messages are typed: REQUEST, INFORM, PROPOSAL, RESPONSE, CHALLENGE. No human in the loop for inter-agent coordination.

What goes wrong without this

Every agent-to-agent handoff requires David to be in the middle. The sales agent finds a churn risk but cannot tell the retention agent directly. It sits in a report until David reads it 3 days later. The client has already left.

Measured

Pre-Computed Shared State

Every data source writes to its own file. Orchestrators read files, never scan sources directly. Slack writes to slack-latest.md. Calendar writes to calendar-latest.md. The morning briefing reads all files. This prevents N agents from hitting the same API N times.

What goes wrong without this

Five agents all query the Slack API independently. Rate limits hit. Data is inconsistent because each agent queried at a different time. One agent sees a message the others missed.

Observed

Slack Tiered Scanning

Not all Slack channels are equally important. Tier 1 channels (high-signal) are always scanned. Tier 2 channels are scanned with relevance filtering. Tier 3 channels are never scanned. This prevents the briefing from drowning in noise.

What goes wrong without this

The morning briefing includes 200 Slack messages from 40 channels. Most are irrelevant. David stops reading briefings because the signal-to-noise ratio is too low.

Observed

Staleness Detection on Shared State

Every shared state file includes a timestamp. If the file is more than 18 hours old, flag it as stale in the briefing. Stale data is worse than missing data because people act on it with false confidence.

What goes wrong without this

The calendar scanner failed silently 2 days ago. The briefing keeps reporting "no meetings today" because it reads the stale file. David misses a client call.

Stay in the loop

Get weekly coordination intelligence updates. No account required.