The canonical agent architecture: A while loop with tools (www.braintrust.dev)

A recent exploration of agent architecture highlights the surprisingly simple core behind many successful AI agents, such as Claude Code and OpenAI’s Agents SDK: a while loop that iteratively calls tools based on a language model’s decisions. This canonical design strips away complexity by focusing on just three components—an LLM, a system prompt, and a small set of well-scoped tools—enabling developers to build flexible, composable, and maintainable agents. This minimal loop structure naturally supports advanced extensions like sub-agents and multi-agent systems, allowing for scalable complexity without convoluted frameworks. Key to this approach is thoughtful tool design and context engineering. Instead of overloading a tool with all possible API parameters, defining narrowly scoped tools with essential inputs tailored to the agent’s objectives reduces cognitive burden on the model and improves reliability. Since tool interactions contribute nearly 80% of an agent’s token context, concise, cleanly formatted tool outputs are as crucial as prompt engineering for effective reasoning. The architecture also encourages building rigorous evaluation pipelines—combining agent tasks, datasets targeting failure modes, and continuous benchmarking—to track progress and prevent regressions as models and prompts evolve. This canonical while-loop architecture exemplifies “the Bitter Lesson” in AI, showing that simplicity and clear modularity outperform complex frameworks. By absorbing complexity into tools and context, while maintaining a straightforward orchestration loop, this pattern offers a robust foundation that adapts gracefully to future model improvements and evolving AI ecosystems.