Planning
Working out a sequence of steps before taking them — the thing agent demos imply models can do, and the evidence says they mostly can't.
Reviewed July 12, 2026Stable
When not to use it
- When you know the sequence. Write it down. This is most cases.
- When ordering errors are expensive. A bad plan reads exactly like a good one until it executes.
- On genuinely novel structure. Unfamiliar constraints are where the obfuscation result bites.
- Autonomously, over long horizons. The evidence doesn't support it, whatever the demo showed.
Reach for something else instead
- A hard-coded workflow — if you can write the steps, this is faster, cheaper, and correct.
- ReAct-style incremental — never needs a plan; decides the next step from feedback.
- Classical planners (PDDL) — solved this in the 1970s. Use the LLM to translate into them.
- Human plan, model execution — the person orders the steps, the model does them.
Read more on the blog
- How AI agents actually work: the loop behind the hypeEveryone is building "agents" and almost nobody explains what makes one. Not the marketing, the actual machinery: the loop, the four components, and the single property that separates an agent from a chatbot with a fancy system prompt.
- Why AI agents fail: the seven failure modesGartner predicts over 40% of agentic AI projects will be canceled by 2027. The failures follow patterns, seven of them. The taxonomy: what breaks, why, which real incident proved it, and which control would have prevented it.
Further reading
- Valmeekam et al. (2023), On the Planning Abilities of Large Language Models: A Critical Investigation — the obfuscation result; the paper agent marketing skips.
- Yao et al. (2023), Tree of Thoughts: Deliberate Problem Solving with Large Language Models — search over candidate steps; better, and expensive.
- Liu et al. (2023), LLM+P: Empowering Large Language Models with Optimal Planning Proficiency — the model translates, a classical planner reasons. It works, and it's unfashionable.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Reading a fluent plan as a good plan. Fluency is free; correctness isn't.
- Testing on conventional tasks. It's seen those orderings. Test where the constraints are unusual.
- Assuming demos generalise. Agent demos are chosen; your task wasn't.
- Ignoring classical planners because they're old. They do the search correctly, which is the part the model can't.
At a glance
FieldAI Agents
What the evidence saysmuch weaker than demos imply
The diagnosticobfuscate the problem; structure-reasoners are invariant, pattern-matchers collapse
Better atvalidating plans than generating them
Best current shapeLLM translates, symbolic planner reasons
DifficultyIntermediate
Often compared with
LLM planning vs. classical planning — one produces fluent plausible sequences, the other does actual search with backtracking. The best current systems use the LLM to translate into the second.
Where this sits
29 concepts come first. Understanding it opens up 1 more.
15Levelsteps in
29Needs firstconcepts
1Opens up0% of 310
1Areastays here
Computed from the prerequisite graph, not assigned. How this works