Skip to content

Stigmergy

Stigmergy is a managed agent fleet platform. You describe what you want built. A Queen agent decomposes the work, spawns Workers, and coordinates them to build it autonomously.

You → Describe your project → stigmergy.yaml
Hivemind (platform)
Hive (Queen + Workers)
Code gets written
  1. Describe: Tell Stigmergy what you want — “Build a SaaS with auth, billing, and a React dashboard”
  2. Review: Stigmergy generates a stigmergy.yaml — the deployment spec for your agent fleet
  3. Deploy: Hit deploy. The platform spawns a Queen, which spawns Workers, which start building
  4. Watch: Monitor progress via dashboard. Every agent action is auditable via Nectar

The top-level container. Contains the Queen, her orders, and all the nests (work streams). One hive per project.

An Opus-tier Claude Code instance with a management prompt. The Queen:

  • Reads orders.yaml — the project spec
  • Decomposes work into parallel tracks (frontend, backend, infra)
  • Writes brood.yaml configs for each work stream
  • Spawns Worker nests
  • Monitors progress via telemetry
  • Intervenes when something goes wrong

The Queen doesn’t write code. It thinks, delegates, and course-corrects — the same workflow a human tech lead uses.

One incubator instance with agents, protocol, and coordination. Each nest handles one track of work (e.g., “frontend” or “backend”). Multiple nests run in parallel within a hive.

Claude Code instances in containers with full dev tools. Workers claim tasks, write code, run tests, and publish completion events. They don’t know they’re in a fleet — they think they’re developers on a team.

Stigmergy automates the exact workflow that built the company. One person (Jocelyn) + Claude Code, coordinating across sessions, monitoring output, intervening when stuck, killing sessions that go nowhere.

Human does this todayQueen does this in Stigmergy
Opens Claude Code, gives it a taskSpawns a Worker with a system prompt
Reads terminal output, decides next stepReads telemetry, adjusts via ACP
”Stop, the tests are failing”Broadcasts halt via ACP events
Runs 3 sessions for frontend/backend/infraHive with 3 Nests
Kills a stuck session, starts freshKills Worker, spawns replacement
Writes CLAUDE.md instructionsWrites system prompts for Workers

The Queen IS the human with a system prompt. Same tools, same decisions, same patterns.

Queen (Opus) → Strategic decisions, task decomposition
Workers (Sonnet) → Complex implementation, architecture
Workers (Haiku) → Repetitive tasks — CSS, tests, migrations
Drones (Haiku) → Review, voting, coordination

The Queen auto-scales model tiers based on task complexity. Simple CSS task → spawn a Haiku worker. Complex API design → Sonnet. The cheapest agent that can do the job.

Queens can spawn sub-Queens for complex projects:

orders.yaml
nests:
frontend:
brood: frontend-brood.yaml
topics: [api.ready, deploy.trigger]
backend:
brood: backend-brood.yaml
topics: [api.ready, db.migrated]
infrastructure:
hive: # Sub-hive with its own Queen
orders: infra-orders.yaml
topics: [deploy.trigger, health.*]

The infrastructure track gets its own Queen, which spawns its own Workers. All connected via ACP topic channels. Recursion depth is configurable.

Workers spin up and down instantly. The meter runs only when agents are thinking:

  • Worker finishes task → publishes completion → container dies → zero cost
  • Queen needs more hands → HTTP call → new Worker in seconds
  • 3 AM, nothing happening → everything down → zero idle burn
  • User pushes code → webhook → Queen wakes → scales as needed

Traditional: ~$0.05/hr per agent (80% idle). Stigmergy: pay only when bees are buzzing.