The Case Against Agent Frameworks

Most agent frameworks force developers to write code-heavy abstractions that obscure what the model actually sees. The author argues that these frameworks often act as "SSH with vibes," creating systems that are difficult to debug, version, or audit. Instead of relying on complex graph-based frameworks, the author suggests building a minimal 'kernel' that treats agents as isolated processes. This approach prioritizes simplicity, allowing non-technical team members to define agent behavior via simple markdown files.

Building a Reliable Agent Runtime

To move from fragile demos to production-ready automation, the author implemented three core engineering principles:

• Event-Driven Architecture: Replace rigid cron jobs with an event-based system. Agents should subscribe to events (e.g., "voice note received," "email arrived") rather than living in a hard-coded graph. This allows for easy fan-out and decoupled system topology.
• Content-Addressed Prompts: To solve the problem of "untraceable" prompts, represent prompts as a list of hashes rather than rendered strings. By hashing individual components (system messages, tool definitions, user input), you gain the ability to diff runs, audit exactly what the model saw, and replay requests against different models for evaluation.
• Strict Type Boundaries: Use typed tool calls and typed events as the primary interface boundaries. This prevents malformed data from propagating through the system. The author notes that roughly 20% of events were initially malformed, and enforcing strict schemas at the kernel level is the only way to make bad actions impossible rather than just unlikely.

Operational Insights

• Build Before You Buy: The author advises against purchasing off-the-shelf agent platforms until you have built a custom solution. This process reveals the specific limitations of current models and helps you understand your actual requirements.
• Observability is Mandatory: Treat agent failures as standard software orchestration problems. An append-only event log is essential for debugging, as it provides a causal link between events, allowing you to trace why an agent produced a specific output.
• Dogfooding: The author emphasizes that developers of agent frameworks must use their own tools. If a framework is too complex to manage your own internal workflows, it is likely too complex for production use.