Agent Frameworks Considered Harmful — Rémi Louf, .txt
The talk details the creation of a robust, event-driven agent runtime system designed to overcome limitations in existing AI frameworks and chat interfaces. The core innovation is moving from string-based prompts and complex graphs to an auditable, content-addressed store where every component (system message, skill description, user question) is hashed and stored separately. This architecture ensures full traceability, allowing for precise diffing between runs and reliable replaying of requests against different models, making the system highly debuggable and scalable.
Key takeaways
-
Event-Driven Architecture (EDA)
15:37
The system relies on agents subscribing to typed events rather than maintaining complex graphs with edges. This simplifies orchestration significantly, allowing for 'funneling' and 'fan out' without requiring code knowledge; users only need to know what events exist in the system.
-
Content-Addressed Prompt Store
20:11
Prompts are no longer rendered strings. Instead, every part (system message, skill description, tool definition, user question) is hashed and stored as an identifier. A prompt is represented as a list of these hashes, enabling precise auditing.
-
Enhanced Observability
The system implements two key boundaries: typed tool calls and typed events. This rigorous typing prevents malformed or non-existent inputs from breaking the pipeline, which was necessary because early failures included rejected/malformed events.