An agent is a struct with state, actions, and tools, running inside a supervised process. See the anatomy of an agent and how each piece connects.

• • Define agents with state, actions, and tools
• • Run agents under OTP supervision
• • Inspect and test agent behavior independently

Tools are typed Actions with input/output contracts. Define a tool, attach it to an agent, and let it call APIs, query databases, or browse the web.

• • Define tools as typed Actions with contracts
• • Attach tools to agents at definition time
• • Validate tool I/O at execution boundaries

Works with OpenAI, Anthropic, Google, Mistral, and local models. Swap providers without changing agent code.

• • Provider abstraction through req_llm
• • Model capability tracking with llm_db
• • Swap models without code changes

Each agent runs in its own BEAM process under OTP supervision. When an agent crashes, its supervisor restarts it in milliseconds with clean state.

• • Process isolation for every agent
• • Automatic crash recovery via OTP supervisors
• • No external orchestrator or health checker needed

Agents communicate through Signals and respond with Actions. Coordination is explicit code you can read, test, and debug.

• • Typed Signals for inter-agent communication
• • Route-to-Action coordination contracts
• • Test coordination logic independently

Built-in telemetry emits events for every agent lifecycle transition. OpenTelemetry integration traces workflows across processes.

• • Telemetry events for agent lifecycle
• • OpenTelemetry trace integration
• • Cross-process workflow tracing

Add one agent to your existing Elixir app. No rewrite, no platform migration. Each package composes without lock-in.

• • Start with one supervised agent
• • Add packages incrementally as needed
• • No lock-in or platform rewrite required