Bounded Autonomy: Engineering AI Agents with Constraints

The Breakthrough

Effective agent performance is achieved by imposing artificial constraints on context and complexity, rather than relying on large context windows or automated internet-based retrieval.

What Actually Worked

• Prioritize Curated Documentation: Replace general internet access with high-quality, domain-specific documentation to prevent the model from absorbing SEO-optimized noise or promotional content.
• Minimize Context Windows: Instead of maximizing context, identify the smallest amount of information required to complete a task. This reduces noise and forces the developer to understand the underlying data structure.
• Use Simple Formats: When generating structured output or documents, prefer simple markup like HTML over complex agentic workflows. In one instance, replacing a complex agentic CV generator with a simple HTML template resulted in a 100x improvement in output quality.
• Build Custom Harnesses: Experiment with building custom memory, compaction, and preprocessing layers rather than relying on black-box agent frameworks. This builds a deeper understanding of the data and improves control over the model.

Context

In high-stakes environments like advertising, where agents generate thousands of assets daily, the tendency is to build increasingly complex systems to manage scale. The author argues that models are naturally verbose and prone to complexity, which leads developers to build temporary, superficial band-aids. By treating the LLM as a flexible database capable of semantic math rather than an autonomous agent, developers can build more robust, predictable systems that prioritize speed and iteration over bloated automation.

Notable Quotes

• "Just because you have the power of the gods that doesn't mean you should use it."
• "The challenge is no longer getting context in but to a certain extent keeping the noise out."
• "Don't automate a job unless you can do it yourself."

Content References

{ "type": "paper", "title": "Attention Is All You Need", "author": "Vaswani et al.", "context": "cited" }