← back
Real AI Agents Need Planning, Not Just Prompting - Yuval Belfer
Takeaway
ReAct agents can't see the forest for the trees — explicit pre-execution planning (form- or code-based) is what makes complex agents reliable.
Summary
- AI21's Yuval Belfer argues ReAct-style step-at-a-time agents lack a lookahead plan — they observe, think, act locally but can't reason about the whole solution trajectory.
- Form-based planners (Magentic-One from Microsoft) and code-based planners (smolagents from HuggingFace) provide explainable, parallelizable plans before execution.
- Even GPT-4.1 still struggles with instruction following from a single mega-prompt — agents need explicit planning as a separate stage.
- Planning matters most for complex tasks requiring parallelization and explainability.
agentsplanningreact
Original description
AI agents that actually deserve the name - do they even exist? Despite the hype, most "agents" today are just LLMs with fancy prompt engineering tricks, lacking true agency capabilities. Here's a deeper issue: it's 2025, and LLMs still struggle with basic instruction following. Weird when one of the first big models was literally called "InstructGPT," right? Benchmarks are saturated but meaningless, and without genuine planning abilities, these systems will keep hitting the same walls. In this session we will go through: - Why conventional agent frameworks like ReAct miss the mark on true agency - How dynamic planning creates agents that actually follow complex instructions - Tips to improve instruction following in any AI system you build