Workflow Architecture

Simple Explanation

A plan becomes a graph of steps. We run ready steps in parallel with limits, retry on errors, and combine results.

Technical Explanation

Planner outputs a DAG. Dispatcher executes ready nodes concurrently (asyncio.gather) under semaphores and budgets; applies retries/circuit breakers; sandboxes code; aggregates outputs; and emits metrics/logs with redaction.

Execution is a DAG with guarded fan-out, retries, and fallbacks.

Key components - Planner output → steps with dependencies, parameters, and stop conditions. - Dispatcher → runs ready nodes concurrently, enforces timeouts and budgets. - Safety rails → retry with backoff, circuit breakers, idempotency keys, redaction. - Sandboxed code → restricted builtins, no network, separate process.

Flow 1) Validate plan (shape + limits). 2) Resolve tools via registry/search. 3) Execute ready steps in parallel (asyncio.gather), apply retries/fallbacks. 4) Aggregate outputs; short-circuit on fatal errors per policy.

Tuning - Concurrency: cap fan-out to avoid runaway spend. - Timeouts: per-step + global; propagate cancellations. - Fallbacks: define alternates for flaky integrations.

Observability - Metrics: success/failure, latency, cache hit rates. - Logs: structured with redaction; debug mode adds more detail safely.