According to LessWrong, an essay breaks down the core architectural patterns for building agentic AI frameworks—essentially, different ways to orchestrate and loop language model API calls to handle complex tasks. The author identifies four main paradigms: sequencing, where LLM calls happen one after another with outputs feeding forward; branching, where multiple calls run in parallel on independent subtasks, the basis for most workflow tools today; looping, where a model repeatedly calls itself in a cycle until it solves the problem, the pattern behind Claude Code and similar assistants; and recursion, where models can delegate to other model instances to manage context and complexity. The essay explores mixing and matching these architectures—swapping single agents for sequences, or combining recursion with looping into multi-agent ecosystems. The practical bottleneck isn't the design itself; it's calibration, error handling, and context management, all of which work because frontier labs fine-tune their models specifically for tool-calling. The author suggests further experimentation could unlock new architectural patterns.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton