Spawn a team. Feel the overhead.
You can spawn a team now. Team lead, three workers, shared task list. It looks like coordination. The agents message each other, claim tasks, report back. It feels like a team.
Then one agent rewrites a file another agent is reading. A third agent gives confident advice based on a hallucination from the first. The "team lead" spends 30% of its tokens just routing — not thinking. You add a fourth worker and everything gets slower.
These aren't bugs in the framework. The frameworks give you plumbing. The hard problems are underneath.
Every multi-agent system you'll ever build is some mixture of two ideas:
The supervisor — One agent owns control flow. It decides who acts, when, with what context. Debuggable. Auditable. A single point of failure and a throughput bottleneck. pentagon.run makes this spatial and visual — a canvas where every agent has a desk.
The swarm — No boss. Agents coordinate through shared state, environment signals, or direct peer-to-peer messaging. Coordination emerges from local rules. The ant colony. claude-peers-mcp is the minimal version: a broker, messages, and nothing else. MiroShark is the extreme: thousands of agents with personalities forming emergent consensus.
Most production systems end up hybrid. A pipeline for structure, a swarm for exploration. A supervisor for synthesis, peers for the actual work.
The load-bearing concepts. Each one grounded in where the research actually is.
- supervisor — One agent to rule them all. When it works, when it breaks, what it costs.
- swarm — No boss, emergent coordination. From ant colonies to a million simulated humans.
- communication — How agents actually talk. Text as bottleneck. KV-cache sharing as frontier.
- memory — Shared state is a lie. Consensus memory, context rot, the consistency models nobody's applying.
Hands-on. Pick one, build something, break something.
- peer-to-peer — claude-peers-mcp. Two sessions, no hierarchy, just messages.
- supervise a team — pentagon.run or Agent Teams. Assign roles, watch the canvas, feel the overhead.
- spawn a swarm — MiroShark. Feed it a document, watch hundreds of agents argue their way to consensus.
The unsolved problems. The conversations worth having.
- the coordination tax — Measure it. When does adding an agent make things worse?
- beyond text — What if agents could share thoughts instead of words?
- your reports are stochastic — You're already an engineering manager. But your reports hallucinate.
Tools for tonight:
| Tool | What it is | Pattern |
|---|---|---|
| claude-peers-mcp | Peer-to-peer messaging between Claude Code sessions | Swarm / stigmergy |
| pentagon.run | Spatial canvas for managing agent teams | Supervisor / mission control |
| MiroShark | Multi-agent simulation engine, hundreds of agents with personalities | Swarm intelligence at scale |
Bring a codebase that needs parallel work, a document worth simulating reactions to, or just curiosity about why your agents keep stepping on each other.
Leave with an intuition for when coordination helps and when it hurts — and the patterns underneath both.
No lectures. Just the structures underneath.