Cambrian Agent SDK

Build Cambrian agents in Python with zero boilerplate.

An agent is a small Python class. The SDK gives it a ReAct loop, tool calling, working memory, and a gRPC server that speaks the Cambrian agent-plane contract — you write run() (or just declare @tool methods) and the kernel does the orchestration.

pip install cambrian-agent-sdk      # or: uv pip install cambrian-agent-sdk

Quickstart

from cambrian_agent_sdk import CognitiveAgent

class AnalystAgent(CognitiveAgent):
    role = "You are a rigorous analytical reasoning engine."
    constraints = ("Ground every claim; say so when the context is insufficient.",)

agent = AnalystAgent(agent_id="analyst_agent", description="Analyses and compares information.")

if __name__ == "__main__":
    agent.serve()   # boots the gRPC agent server; the kernel connects and dispatches tasks

The kernel spawns your agent, hands it tasks over the agent plane, and your run()/think() loop can call memory (memory_query), tools (tool_call — kernel-owned or your own @tool methods), and delegate (yield_subgoal) to the planner.

Agent shapes

• CognitiveAgent — LLM reasoning: run() → think() ReAct loop, memory, and @tool actions.
• DeterministicAgent — a scripted cell: typed run() only, no reasoning (a fast tool).
• DaemonAgent — a persistent background producer on the SignalStream contract.

Declare tools with the @tool decorator (auto-injected into the agent's action menu) and capabilities with @capability. See example-agents/ in the repo.

Versioning

The SDK's minor version tracks the agent-plane proto contract it targets. Pin it like any dependency; the vendored gRPC/protobuf gencode is shipped inside the wheel, so you don't run protoc yourself. The runtime requires protobuf>=6.31,<7.

License

Apache-2.0. See LICENSE.