References · Further reading
01 / 03
Talk · AI Engineer World's Fair 2025
How We Build Effective Agents
- The minimal agent is just three things — environment + tools + system prompt. Everything else (memory, planners, sub-agents) is downstream optimisation — add it only when you've felt the pain it solves.
- Three rules for builders: don't build agents for everything; keep it simple; think like your agent — get inside its head before debugging the prompt.
arXiv · 2210.03629
π
ReAct.
Paper · ICLR 2023
ReAct: Synergizing Reasoning and Acting in Language Models
- Interleaving reasoning traces with tool actions beats either alone on reasoning, QA, and decision-making benchmarks — and it makes the model's behaviour interpretable along the way.
- This is the foundational pattern every modern "agent framework" is implementing. Strip a 2026 agent product down to its core and this loop is what's left.
- Published as an arXiv preprint in October 2022 — before ChatGPT shipped. The agentic loop was a research idea two months before the "agent" buzzword existed, and it has held up across every model generation since.
Talk · AI Engineer World's Fair 2025 · Best Speaker
Pelicans on Bicycles
- "Tools + reasoning is the most powerful technique in AI engineering right now."
- The one-line thesis for everything that follows in this webinar.