Why Agent Runtimes Need Events and Logs

AI Engineer20:29
0 comments · 0 votesOpen discussion

Everyone can read the discussion. Sign in to comment, reply, vote, or report abuse.

Sign in to join the discussion

    Video summary

    Rémi Louf describes building unattended agents for market research, voice-note processing and daily briefings. Simple schedules worked initially, but duplicate runs, lost notes and untracked prompt changes exposed ordinary distributed-systems problems that a prompt alone could not solve.

    Rémi Louf replaces a fixed workflow graph with agents that subscribe to typed events. An append-only log preserves every step and causal link, while queues, retries and structured outputs make failures observable and prevent invalid actions from crossing system boundaries.

    A content-addressed prompt store records the exact system prompt, skills, tools and user message that reached each model. That representation supports diffs, replay with different models, compaction, cache management and auditing, while keeping the agent definitions simple enough for non-engineers to contribute.

    Watch on YouTube