Rémi Louf: Agent Frameworks Are "Harmful"

Rémi Louf, CEO of .txt, presented a compelling case against current agent frameworks at the AI Engineer World's Fair. Louf, whose company builds infrastructure for reliable agents, argued that the prevailing frameworks are "harmful" because they often fall short of delivering truly autonomous agents. He shared his personal journey over two weeks in January, spurred by significant improvements in AI agents, to understand their capabilities and identify better building primitives.

The Promise of Autonomous Agents

Louf began by illustrating his ideal vision of an autonomous morning routine: a "morning briefing with my coffee." This would involve agents automatically browsing market news, reviewing his CRM and Jira, and processing long voice notes recorded during his walks. However, he found that current AI tools, while capable of coding and other tasks, often require continuous user input, likening it to "having a robot mower that you still have to stay on." He critiqued the transitional nature of current solutions, such as apps that feel like "SSH with vibes," where users are still actively managing the agent.

From Frustration to Building

Frustrated by the limitations, Louf decided to build his own solution, starting with the "simplest thing that could possibly work." He found that while coding frameworks exist, the real challenge lay in prompt editing. He discovered that using YAML frontmatter, as opposed to pure code, made implementing agents easier, allowing for versioning, diffing, and PR reviews. The system also leverages cron jobs for scheduling and emits events for new data like voice notes or emails, enabling agents to react dynamically.

Events Over Graphs

Louf emphasized the importance of an event-driven architecture, contrasting it with graph-based approaches. "No edges to maintain, agents subscribe to events," he stated. This event-centric model allows for fan-in and fan-out capabilities and simplifies agent creation, requiring only a file drop and no complex coding. He demonstrated this with a voice note processing agent that accepts a voice note, transcribes it, and emits a new "voice_note.processed" event, which in turn triggers a daily brief agent.

The Runtime Solution

Louf detailed the failures he encountered, such as duplicate Slack messages and vanished voice notes, which led to the development of a robust runtime. Key components include an immutable log for traceability, a proper queue system, and a content-addressed system (akin to Git) for managing prompts and artifacts. This system allows for precise tracking of what went into the model, ensuring auditability and making compaction easier. He also highlighted the ability to perform diffs between runs and replay requests with different models or parameters.

Key Takeaways for Agent Development

Louf concluded with several key lessons for building effective AI agents:

• Well-executed background agents are "magical," automating tasks and processing information seamlessly.
• The difficulties encountered are primarily "good old engineering problems" related to orchestration.
• Open-source models are readily available and "good enough" for many applications, even running locally.
• The infrastructure category for agents is "unsettled," and Louf advises building before buying to understand specific needs and limitations.
• Developers building agent frameworks should "eat your own dog food."
• The field is moving rapidly, necessitating active engagement and experimentation.

Louf stressed the importance of immersing oneself in the technology, as his two-week experiment fundamentally shifted the trajectory of his company. He encouraged others to "steal this code" from his GitHub repository, emphasizing that it's not a product they intend to sell.