VibeCodingNights/agent-orchestration
your agents work alone. even when there are five of them
Platform-specific configuration:
{
"mcpServers": {
"agent-orchestration": {
"command": "npx",
"args": [
"-y",
"agent-orchestration"
]
}
}
}Add the config above to .claude/settings.json under the mcpServers key.
your agents work alone. even when there are five of them.
---
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. they're the fundamental problems of distributed systems, consensus, and coordination that computer scientists have been working on for fifty years. 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.
Loading reviews...