Agent Frameworks Considered Harmful: Building Robust AI Systems from First Principles

Rémi Louf points out that traditional Cron jobs, while commonly used, are fundamentally inadequate for complex event-driven agent systems. These jobs are designed for time-based execution, scheduling tasks to run at specific intervals. However, AI agents need to respond dynamically to a variety of unpredictable system events, such as file uploads, incoming emails, or updates to pull requests. Relying on Cron jobs creates significant gaps in system observability and state management, as they cannot account for events occurring between scheduled runs or the specific triggers of an action.

The proposed solution shifts away from time-based scheduling towards a system-wide event-based reactivity. In this paradigm, new data inputs, like dropping a voice note into the system, automatically emit an event. Agents are then designed to react specifically to these events, rather than periodically polling for changes. Each agent is built with clearly defined inputs and outputs, and the enforcement of structured outputs allows agents to reliably chain tasks together, creating a robust and predictable workflow.

Louf criticizes agent frameworks that advocate for graph-based logic, asserting that such approaches introduce unnecessary complexity. These frameworks often require developers to manually map out intricate execution graphs, which can be difficult to maintain and scale. Instead, an event-based subscription model eliminates the need to maintain static edges in a graph. The system's topology emerges dynamically from the event logs, simplifying orchestration and making the system more flexible and adaptable to changes.

Early in the development of their agent system, .txt encountered several critical failures that directly informed their architectural choices. These included duplicate Slack messages for daily briefs, the complete disappearance of a voice note, and market briefs turning into "garbage" due to unversioned prompt edits. These real-world problems highlighted the limitations of the initial setup and underscored the necessity of a robust runtime capable of handling such issues systematically.

Each failure mode encountered during development was systematically converted into a core architectural component of the runtime. The data loss from vanished voice notes led to the creation of an append-only event log, ensuring that no information is ever discarded. Duplicate task executions necessitated a proper task queue capable of counting and managing attempts. The problem of untraceable prompt edits resulted in a content-addressed storage system, providing immutable versioning for all prompts. Together, these components form a comprehensive 'system's memory' where every event is causally linked and permanently recorded, creating a resilient and observable system.

Louf describes agents as isolated processes within a system, drawing an analogy to an operating system kernel. The system is designed to schedule and isolate these agents, abstracting away their internal workings. This isolation allows non-technical users to define agent behaviors using simple markdown, without needing deep technical knowledge of the underlying system. The system journals agent definitions within its logs, maintaining a clear record of their configurations and actions.

A key design principle of the system is to make "bad actions impossible, not just unlikely," a goal that emerged from addressing high error rates in structured outputs from earlier models. This is achieved through the enforcement of strict type boundaries. Specifically, there are two critical boundaries: typed tool calls, which ensure agents interact with external tools predictably, and typed events between agents, which guarantee that data passed between agents conforms to predefined schemas.

Currently, .txt operates 20 internal agents, contributed by both technical and non-technical staff, demonstrating the accessibility afforded by markdown-based agent definitions. These background agents automate critical tasks, such as generating daily briefings, which Rémi Louf asserts are "really magical" and superior to manual processes. He emphasizes that the primary challenge in this domain lies in engineering orchestration rather than specific AI complexities, underscoring the importance of robust system design.

Rémi Louf advises organizations to "definitely start building before you buy" when it comes to AI infrastructure. He highlights that the infrastructure category for AI agents is still nascent and unsettled, making early purchasing risky. For small companies or tech CEOs, building internal tools first can be an advantage, allowing them to understand their specific needs and limitations without relying on external vendors. He also urges framework developers to "eat their own dog food," meaning they should build practical applications with their own tools to ensure their utility and address real-world problems.