Stop using agent frameworks: Build event-driven pipelines instead

Rémi Louf, CEO of txt, took two weeks to build a personal agent system for his morning routine. He started with agent frameworks but found them too code-heavy and rigid. Instead, he designed a YAML-based, event-driven system using markdown files and cron jobs.

The system uses structured outputs for reliability, content-addressed prompts for debugging, and typed events for inter-agent communication. After initial failures—lost voice notes, duplicate Slack posts, unversioned prompt changes—he added logging, queues, and versioning. The result: 20 agents running in production, contributed by non-technical team members.

Louf argues that agent orchestration is not a novel problem—it's "good old distributed systems." He claims event-driven architecture (not graphs or frameworks) is the right primitive. His key insight: structured outputs are non-negotiable, and infropic models often fail on tool calls, which is why his company's product became essential.

The talk is part blueprint, part manifesto. It warns against premature framework adoption and urges teams to build before buying. Louf replaced all third-party APIs with open-source models, claiming they are "good enough" for background agents. The evidence is strong: he shipped a working system used by a real company.