← back
Keynote: Why people think "agent" is a buzzword but it isn't
Takeaway
Agents aren't a buzzword — they're just hard because step failure compounds; the unlock is models with stronger planning plus better tools and memory.
Summary
- Chip Huyen uses Russell & Norvig's classic definition (anything that perceives and acts on an environment) to ground agents — ChatGPT, SWE-agent, coding agents are all examples.
- Curse of complexity: even at 2% per-step failure, 10 steps yields ~18% failure, 100 steps near-worthless; most current agents only reliably handle ~5-step tasks.
- Newer reasoning models (DeepSeek R1, Gemini 2.0 Flash Thinking, o1-preview) materially shift planning curves upward on her synthetic planning benchmark.
- Other reasons agents are hard: poor tool design, weak or wrong tool inventories, and need for memory beyond what the context window can hold.
agentsplanningreasoning
Original description
Chip Huyen's keynote for the AIE Summit Online track