A live workshop (with significant technical difficulties) exploring an event-sourced architecture for building AI agent harnesses. The core idea is that every action in an agent loop — LLM requests, tool calls, streaming chunks, errors — becomes an immutable event in an append-only stream. A stream processor pattern (reduce function + side-effect hook) consumes these events to derive state and trigger actions. Key properties of the proposed system include: full debuggability since all state is derivable from the event log, distributed extensibility where plugins written in any language can subscribe to and append events, edge deployability so every agent has a public URL, and a 'dynamic worker' feature where you can deploy a processor simply by appending an event containing its JavaScript source code. The workshop demos a proof-of-concept service at events.iterate.com with SSE-based streaming, idempotency keys, circuit-breaker pausing, scheduled events, and push/pull subscriptions.
Sort: