Agents
Source: https://huyenchip.com/2025/01/07/agents.html Author: Chip Huyen Date: 2025-01-10
Summary
Chip Huyen’s comprehensive overview of AI agents — architecture, tool use, planning, memory, and failure modes. One of the best single-page references on agents as of early 2025.
Key Claims
- Agent = model + tools + memory + planning + action. Removing any one changes what it can do.
- Tool taxonomy: read tools (retrieval), write tools (actions), compute tools (code execution). Most failures come from write tools.
- Planning strategies: ReAct (reason-act), ToT (tree of thought), MCTS (monte carlo tree search). Most production agents use ReAct due to simplicity.
- Memory taxonomy: in-context (working memory), external (RAG/files), parametric (model weights). Most agents only use first two.
- Failure modes: hallucination, tool misuse, loop traps (agent gets stuck), context overflow.
- Evaluation challenge: agent success is often binary and end-to-end — hard to give partial credit or debug intermediate steps.
Entities
- Chip Huyen — ML systems author; huyenchip.com
Concepts
- Coding Agents — primary application domain
- Agent Memory — taxonomy from this source
- Context Engineering — context overflow as failure mode
- Test-Time Compute — ToT/MCTS as test-time compute strategies