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.
Architecture
Data Before Design
Collect data about a blind spot for at least 2 weeks before designing a solution. The sales agent ran in shadow mode (reporting only) for 4 weeks before getting write access. This prevents building elaborate solutions for problems that do not exist.
What goes wrong without this
You build an elaborate lead scoring system. After launch, you discover that 90% of leads come through one channel and a simple FIFO queue would have worked. Two weeks of build time wasted.
Earn Complexity
Validate the current agent stack before adding new agents or tools. Every new agent must justify its existence with a specific job that no existing agent can do. "It would be cool to have an agent for X" is not justification.
What goes wrong without this
The team has 30 agents. 8 of them have overlapping responsibilities. 5 of them run but nobody reads their output. The maintenance burden exceeds the value. Debugging takes hours because nobody knows which agent is responsible for what.
File-Based State Is Authoritative
When file-based state (shared state files, config files) and agent memory conflict, the file wins. Agents must load canonical files before acting on remembered context. Memory supplements. It never overrides.
What goes wrong without this
An agent remembers that Client X is on a $5K/month plan. But the config file was updated to $8K/month last week. The agent reports inaccurate revenue numbers for 2 weeks until someone notices.
Stay in the loop
Get weekly coordination intelligence updates. No account required.