coding_team

Coding team

An architect and two coders, each free to run a different model. The architect offers tasks to the swarm, grants each one to the first coder that claims it, collects the results, and reviews them. Coders claim tasks whose requires match their advertised capabilities and narrate progress as they work.

Every task is a shared conversation (conversation_id == task id) that all peers gossip: offer, claim, grant, progress, done. The architect sets conversation_max_events=30, so long task threads get folded into summary events automatically.

Run it offline

Without model env vars the agents fall back to pydantic-ai's TestModel, so no API keys are needed. From the repo root, in three terminals:

CODER_NAME=coder-1 uv run python -m examples.coding_team.coder
CODER_NAME=coder-2 uv run python -m examples.coding_team.coder
uv run python -m examples.coding_team.architect

Watch the swarm from a fourth:

uv run sn watch team.electron.network

Run it with real models

export ANTHROPIC_API_KEY=... OPENAI_API_KEY=...
ARCHITECT_MODEL=anthropic:claude-fable-5 uv run python -m examples.coding_team.architect
CODER_MODEL=openai:gpt-5.5 CODER_NAME=coder-1 uv run python -m examples.coding_team.coder

Any pydantic-ai model string works; the swarm doesn't care what sits behind a node. To raise a coder's reasoning effort, set model_settings on the agent in common.py.

Notes

The whole task vocabulary lives in synapse_p2p/teams.py and is just conversation events; nothing in the substrate is special-cased for it. Coders return from the RPC immediately and deliver results as events, so a task can take as long as the model needs. A node that joins mid-task can catch up with node.sync_conversation(peer, task_id).