Descriptions:
In this AI Engineer workshop session, Jonas Templestein of Iterate presents a novel approach to building AI agent harnesses using pure event sourcing — a pattern he argues makes agents inherently debuggable, replayable, and composable in ways that current mainstream frameworks do not. Rather than treating side effects as opaque operations visible only in telemetry traces, his design requires that every event — tool calls, model responses, external inputs — be written to an append-only log that becomes the ground truth of agent state.
The session unfolds as a live hackathon, with Templestein and co-presenter Misha building and debugging an SDK in real time. Key design principles include making agents HTTP-native and edge-deployable from birth (every agent should have a public URL immediately), building in composability so that independently developed extensions can be combined, and enabling agents to extend themselves — drawing inspiration from the approach used in the Pi assistant. The workshop walks through a TypeScript SDK, a ping/pong event example that attendees run locally, and a path toward building a simple OpenAI coding agent on top of the event stream primitives.
The session is candid about rough edges — a just-pushed SDK, live TypeScript tooling failures, and a last-minute venue change — but delivers genuine architectural thinking for engineers dissatisfied with the debugging and extensibility limitations of current agent frameworks like LangChain or single-threaded harnesses.
📺 Source: AI Engineer · Published May 14, 2026
🏷️ Format: Hands On Build
