Agents

Chip Huyen published a blog post titled "Agents," adapted from the Agents section of her book AI Engineering (2025) with minor edits. The post provides a framework for understanding AI agents, noting that the field lacks established theoretical frameworks for defining, developing, and evaluating them.

The post defines an agent as anything that can perceive its environment and act upon it, referencing the definition from Artificial Intelligence: A Modern Approach (1995). It states that an agent is characterized by its environment and the set of actions it can perform, with tools augmenting those actions. Examples of agents include ChatGPT, which can search the web, execute Python code, and generate images, and RAG systems, where retrievers and SQL executors are the tools.

The post discusses two aspects that determine an agent's capabilities: tools and planning. Tools are categorized into knowledge augmentation, capability extension, and write actions. Huyen cautions that more tools increase capabilities but also make it harder to understand and utilize them well. She also highlights that agents require more powerful models due to compound mistakes and higher stakes, citing that a model with 95% accuracy per step drops to 60% over 10 steps and 0.6% over 100 steps.

The post addresses planning, noting that effective planning requires understanding the task, considering options, and choosing the most promising one. It recommends decoupling planning from execution to avoid wasteful runs, suggesting validation through heuristics like eliminating plans with invalid actions or too many steps.

The post includes a note that just before the book came out, Anthropic published a blog post on Building effective agents (Dec 2024), and Huyen is glad to see conceptual alignment, though Anthropic's post focuses on isolated patterns while hers covers why and how things work, with more focus on planning, tool selection, and failure modes.