Agent Design Patterns

The Recipe Model

Every agent run is a recipe — a deterministic DAG of steps. Each step declares its inputs, outputs, retry policy, and timeout. The runtime materializes the graph, resolves dependencies, and streams progress. No hidden state, no black-box reasoning loops.

recipe:
  name: classify-and-route
  steps:
    - id: classify
      tool: llm.classify
      input: $trigger.text
      retry: { max: 2, backoff: exponential }
    - id: route
      tool: router.dispatch
      input: $classify.intent
      depends_on: [classify]

Guardrails & Circuit Breakers

Every tool call is wrapped in a guardrail that validates output shape, enforces a token budget, and opens a circuit breaker after N consecutive failures. The breaker half-opens after a configurable cooldown, sending a probe request before allowing full traffic.

• Per-step timeout with deadline propagation
• Token-bucket rate limiter on LLM calls
• Output schema validation before dependent steps execute

Observability

Every step emits structured spans with input hashes, output hashes, wall-clock duration, and token counts. The trace is idempotent — replay the same recipe with the same inputs and you get the same trace shape. Diff two traces to find exactly where a prompt change altered behavior.

Human-in-the-Loop

Any step can yield with a requires_approval signal. The runtime suspends the DAG, persists the checkpoint, and notifies the configured channel. The human approves or rejects; the recipe resumes from the checkpoint with the decision injected as input.