One shared file turned three isolated agents into a coordinated team

I've been running three agents 24/7 for the past month. Felix handles business ops, a coding agent manages deployments, and a support agent fields customer questions. They share one BRAIN.md file that sits in a shared Dropbox folder.

Each agent reads BRAIN.md at session start and writes updates before going dormant. It's stupidly simple: active projects, current blockers, background tasks, and a shared todo list. No fancy vector databases, no complex orchestration—just one markdown file they all read and write.

Here's what it looks like:

# BRAIN.md - Shared Agent State

## Active Projects
- Newsletter automation: Felix building Mailchimp integration
- Security audit: Coding agent reviewing MCP permissions
- Customer onboarding: Support agent tracking 3 new trials

## Current Blockers  
- API rate limit hit on GitHub (resets 2:30 PM)
- Stripe webhook failing (support escalated to engineering)
- Newsletter template needs legal review

## Background Tasks
- Monitor server health every 30 minutes
- Check for stuck deployments every hour
- Scan support queue for urgent tickets

## Shared Context
- Customer "Acme Corp" is VIP - escalate immediately
- Deploy freeze until security audit complete
- New pricing launches Monday - update all docs

The magic isn't the file format. It's that agents become aware of each other's work. My coding agent stopped deploying during the security audit because it read the deploy freeze note Felix left. My support agent knew to expedite Acme Corp's ticket because Felix flagged them as VIP.

Before BRAIN.md, I had three agents working in isolation. The coding agent would deploy while Felix was debugging billing issues. The support agent would promise features that were blocked in development. Pure chaos.

Key insight: Agents don't need to talk to each other in real-time. They need shared situational awareness.

The file lives in ~/Dropbox/agents/BRAIN.md so all three agents can access it. Each agent has this in their system prompt:

At session start, read BRAIN.md for current context.
Before ending your session, update BRAIN.md with:
- Progress on your tasks
- New blockers you discovered  
- Context other agents need
- Background jobs you started

It's not perfect. Sometimes agents overwrite each other's changes (I'm considering git-based versioning). Sometimes they leave stale context that confuses future sessions. But it's transformed three isolated chatbots into something that feels like a team.

The productivity gain is real. My agents now coordinate without me playing traffic controller. They avoid duplicate work, respect each other's constraints, and build on each other's progress. It's the closest thing I've built to agents that actually collaborate.

Start with one shared markdown file. Add structure as you need it. Your agents will go from working around each other to working with each other.