What it is
Planning is working out, before you start, which steps will get you from where you are to where you want to be, and in what order. For an AI, it means taking a goal such as "organise the files in this folder" or "find me a hotel near the conference under $200 a night" and turning it into a list of smaller jobs it can actually do, one after another.
You already plan all the time. A recipe is a plan: preheat the oven first, because it takes ten minutes to warm up, and don't add the eggs until the butter has cooled. A trip itinerary is a plan too: book the flight before the hotel, because the hotel depends on which day you land, and leave two hours at the airport, because missing the flight ruins everything after it. The skill isn't writing a list. It's getting the order right, noticing which steps depend on which, and knowing what you'll do if one of them goes wrong.
That last part matters as much as the list. Real plans meet surprises. The shop is out of eggs; the flight is cancelled. A good planner doesn't abandon the goal. They change the plan: buy eggs elsewhere, take the train. AI researchers call this replanning, and it's the difference between a plan that looks good on paper and one that gets the job done.
For a chatbot answering a single question, planning happens quietly inside the reply, if at all. For an agent, an AI that takes actions on its own such as searching the web, editing files, or filling in forms, planning is central. An agent working through twenty steps needs to know what the twenty steps are, which comes first, and what to do when step seven fails.
There are two broad ways an AI agent can plan:
- Plan first, then execute. The AI writes out the whole plan up front, then works through it. This is predictable and easy for a person to check before anything happens. Its weakness is rigidity: the plan was written before the agent knew what it would find, so it can go stale.
- Decide one step at a time. The AI looks at the situation, picks the next action, sees the result, and decides again. This is the pattern behind the ReAct loop. It adapts well to surprises, but it can wander, repeat itself, or lose sight of the overall goal, because there's no map, only the next turn.
Most capable agents today mix the two. They sketch a plan, carry out a step, check the result, and revise the plan when reality disagrees with it. That's exactly what this element's popup describes: outline, execute, revise.
Two more ideas complete the picture:
- Checkpoints. A checkpoint is a moment in the plan where the work stops and someone, or something, checks it before going further. "Show me the guest list before sending any invitations" is a checkpoint. They catch mistakes while they're still cheap to fix.
- Verification. Somebody has to ask whether each step actually worked, and whether the plan as a whole makes sense. A plan with ten steps where step four is impossible is not a good plan, however tidy it looks.
That brings us to the central surprise about AI and planning. A language model is extraordinarily good at producing text that looks like a plan. It has read millions of recipes, project schedules, and itineraries, and it knows exactly what a plan should sound like. What it can't always do reliably is check that the plan works: that the budget adds up, that nothing is booked twice, that step three doesn't need something only step five produces. The result can be a plan that is plausible but unworkable, and it is written with the same confidence as a good one.
Why it matters
It's what turns a chatbot into something that gets work done. Answering a question takes one step. Booking a trip, cleaning up a spreadsheet, researching a market, or fixing a bug in a website takes dozens, and they have to happen in a sensible order. Without planning, an AI can only react. With it, the AI can take on a whole job. Every agent feature you'll meet, from research modes to coding agents to computer use, rests on some form of planning.
A visible plan is your best chance to steer. When an AI shows you its plan before it starts, you can spot the wrong assumption, the missing step, or the step you didn't want at all, before any time or money has been spent. Correcting a plan takes seconds. Undoing twenty actions built on a bad assumption can take an afternoon. This is why the most useful agent products now have a "show me the plan first" option, and why the research tools that let you edit their plan tend to produce more useful reports.
Plausible plans are a special kind of risk. An obviously wrong answer is easy to catch. A plan that looks professional, with numbered steps, sensible headings, and confident timings, invites trust. But as the examples below show, AI plans have repeatedly failed on the unglamorous details: a budget that doesn't add up, a constraint quietly ignored, an action that isn't allowed. The more authority the AI has to act on its own plan, the more those details matter. This is where human in the loop checks earn their keep.
It explains why agents behave the way they do. Once you understand planning, a lot of agent behaviour makes sense. An agent that keeps repeating the same search is stuck without a plan. An agent that ploughs on after an error is following a plan too rigidly. An agent that asks "shall I go ahead?" is at a checkpoint. Knowing the vocabulary helps you give better instructions and diagnose what went wrong.
It's where the research frontier is. Planning is one of the clearest tests of whether AI can reason, not just recall. The newer reasoning models, those that use extended thinking, were built partly to do better on exactly these multi-step problems, and they do. But the gap between "better" and "reliable" is still wide enough that the best-designed systems pair the AI with a separate checker. Understanding that design will help you judge the claims made about new agent products.
Real-world examples
Shakey the robot and STRIPS (1966 to 1972)
Planning is one of the oldest ideas in artificial intelligence. From 1966 to 1972, researchers at the Stanford Research Institute (now SRI International) in California built Shakey, which the IEEE, the engineering body, recognises as the world's first mobile intelligent robot. Shakey was a wobbly wheeled cabinet with a camera, and SRI says it could perform tasks that needed planning, route-finding, and rearranging simple objects such as boxes in a set of rooms.
To make Shakey plan, Richard Fikes and Nils Nilsson created STRIPS, published in the journal Artificial Intelligence in 1971. STRIPS described every action by three things: what must be true before you can do it, what stops being true afterwards, and what becomes true. "Push the box through the door" needs the robot to be next to the box and the box next to the door; afterwards, the box is in the other room. With actions written this way, a computer can search for a sequence of steps that leads from the current situation to the goal. That simple format became the base for most of the planning languages used today. The same team also produced A*, a route-finding method that, in a later form, runs inside sat-nav systems and video games.
What makes Shakey relevant to modern agents is what happened after the plan was made. The IEEE milestone citation, dedicated in 2017, credits Shakey with being able to "create plans, recover from errors in plan execution". A companion program called PLANEX watched the plan being carried out, spotted when things went wrong, and adapted the plan to what was really happening, which the IEEE describes as groundbreaking for a physical robot. Outline, execute, revise: the loop in this element's popup was running on a robot more than fifty years ago.
PlanBench and "Mystery Blocksworld" (2023 to 2024)
When chatbots got good at sounding like planners, a team at Arizona State University led by Subbarao Kambhampati set out to test whether they really were. Their tests, gathered into a benchmark called PlanBench, used planning puzzles from the research field, including Blocksworld, the classic job of restacking toy blocks into a target arrangement using a few allowed moves.
In a paper presented at the NeurIPS conference in 2023, they reported that the best model they tested, GPT-4, produced a valid plan without help in only about 12 percent of cases on average across their test domains. Then they tried a clever twist. In "Mystery Blocksworld", the puzzle is exactly the same, but the names of the actions and objects are replaced with meaningless words or random strings. A real planner doesn't care what things are called; the rules are identical. A system that is mostly recalling familiar-looking plans should fall apart.
It fell apart. In a follow-up paper in September 2024, the team reported that on 600 ordinary Blocksworld problems GPT-4 solved 34.6 percent, but on the disguised versions it solved just 0.16 percent, a single problem. OpenAI's new reasoning model, o1-preview, did far better: 97.8 percent on ordinary Blocksworld and 52.8 percent on the disguised version. But its accuracy fell to about 24 percent on longer problems needing 20 to 40 steps, and when given problems that had no solution at all, it correctly said so only 27 percent of the time; in 54 percent of cases it confidently wrote out a full plan for a puzzle that could not be solved. The researchers also pointed out that a conventional planning program solves these puzzles perfectly for almost nothing.
Their conclusion, set out in a position paper at the ICML conference in 2024 with the blunt title "LLMs Can't Plan, But Can Help Planning in LLM-Modulo Frameworks", was not that language models are useless for planning. It was that they shouldn't be trusted to plan and check their own work alone. They work best as idea generators alongside a separate checker that verifies each plan and sends back what's wrong. Models have improved since these papers, so treat the exact numbers as a snapshot. The design lesson has held up.
TravelPlanner (February 2024)
Blocks are abstract. Travel is not. In February 2024, researchers led by Jian Xie and Yu Su released TravelPlanner, a benchmark of 1,225 realistic trip requests, later presented at the ICML conference in 2024. Each request asked an AI agent to plan a trip of three, five, or seven days, across one to three cities, using search tools with nearly four million records of flights, restaurants, attractions, and accommodation. Plans had to respect the traveller's own conditions, such as a budget, a room type, a cuisine, or "no flights", and common sense: no visiting the same restaurant twice, no being in two cities at once, and no invented details.
The best agent they tested, GPT-4-Turbo, produced a plan meeting every requirement for just 0.6 percent of the test trips, 6 out of 1,000. Even when all the information was collected in advance, so the model only had to plan, the best pass rate was 4.4 percent. The researchers found agents getting stuck repeating actions, mixing up one piece of information with another, and, most tellingly, satisfying most constraints but rarely all of them at once. The plans looked like itineraries. They just didn't hold together.
The benchmark then became a proving ground for better designs. In April 2024, a team led by Yilun Hao posted a paper in which the language model translated each trip request into a formal puzzle and handed it to a specialised solver program that guarantees every constraint is met. They reported a 93.9 percent success rate. That is the LLM-Modulo idea in action: let the language model understand the messy human request, and let a checker make sure the plan actually works. These figures are from 2024; newer models score higher, but the gap between "looks right" and "is right" is the point.
Plan-and-Solve prompting (2023)
Planning doesn't need an agent. In May 2023, Lei Wang and colleagues published "Plan-and-Solve Prompting", presented at the ACL conference that year. They had noticed that when you ask a model to "think step by step" (see chain of thought), it often skips steps or misreads the problem. Their fix was a different opening line: "Let's first understand the problem and devise a plan to solve the problem. Then, let's carry out the plan to solve the problem step by step."
Tested with GPT-3 across ten sets of maths, common-sense, and reasoning questions, this plan-first instruction consistently beat plain "step by step", and on maths problems came close to prompts that included eight worked examples. The practical lesson for everyday users is simple: asking an AI to plan before it answers is a small change to your prompt that often makes a real difference.
"Plan mode" in coding agents (at the time of writing)
Software developers were among the first to hand AI agents real work, and the tools they use have made planning a visible, separate step. Claude Code, Anthropic's coding agent, has a "plan mode": at the time of writing, Anthropic's documentation says that in this mode Claude reads files and proposes a plan but makes no changes until the user approves it. The user can open the plan and edit it directly before the agent goes ahead. Other coding agents offer similar plan-and-approve options under various names. Features change quickly, so check each tool's current documentation.
What's instructive is Anthropic's own advice on when to use it. Its best-practice guide recommends exploring first, then planning, then writing code, because an agent that jumps straight in can solve the wrong problem. But it also warns that planning adds overhead, and says that for small, clear changes like fixing a typo you should just ask for the change. Its rule of thumb: "If you could describe the diff in one sentence, skip the plan." In plain terms, plan when the job is big, unfamiliar, or touches many things at once; don't bother for a one-liner. That rule works well outside programming too.
Try it yourself
You'll practise the most useful planning habit there is: make the AI show its plan, check it, approve a better version, and only then let it do the work, step by step. Any chatbot will do. Allow about thirty minutes. The example is a small event, but you can swap in a home project such as clearing out a garage or repainting a bedroom.
- Ask for the plan only. Start a new chat and paste this, changing the details to suit your life: “
I'm hosting a backyard 10th birthday party for my daughter in 3 weeks. About 15 children and 10 adults. Budget: $300 total. It might rain. Two guests have nut allergies. I can spend about 4 hours a week on it. Do NOT start doing the work yet. First, write a numbered plan: the steps in order, a rough date for each, the cost of each step, and which steps depend on others. Mark any step where you need a decision from me with "DECISION". List your assumptions at the end.“ - Read the plan like an inspector, not a guest. Don't be won over by the neat formatting. Check it against your real constraints. Do the costs add up to $300 or less? Is there a rain plan, or just a hopeful note? Are the nut allergies handled in the food, the cake, and the party bags? Is anything scheduled in the wrong order, such as sending invitations before the date is fixed? Does the weekly work fit in four hours? Write down every problem you find.
- Ask the AI to critique its own plan. Paste: “
Before we change anything, review your plan as a sceptical professional event planner. Check the arithmetic, the order of steps, the time needed each week, and every constraint I gave you. List each problem you find. Don't rewrite the plan yet.“ Compare its list with yours. It will often catch some problems you missed and miss some you caught. That's normal, and it's why both checks matter (see self-critique). - Approve a revised plan, with checkpoints. Paste your own list of problems and add: “
Here are the problems I found: [paste your list]. Rewrite the plan to fix all of these and the ones you found. Add a CHECKPOINT after any step where I should confirm before moving on (for example, before spending money or sending messages). Show the running total of costs after each step.“ Read it again. When you're happy, reply with a clear approval, such as "Approved. We'll work through it one step at a time." - Execute one step at a time. Ask for only the first step to be done: “
Do step 1 only. Show me the result, then stop and wait for me.“ For a party, a step might be drafting the invitation text or a shopping list. Check the result, then say "Step 1 done. Do step 2 only." Carry on. - Practise replanning. Partway through, change something, the way real life does: “
Change of plan: the budget is now $220, and three more children are coming. Update the remaining steps only. Tell me what you changed and why, and flag anything that no longer works.“ Check whether it kept the finished steps intact and whether the new numbers really add up. - Keep your rule. Write yourself one line from what you saw. For example: "For anything with a budget, a date, or more than five steps, I ask for the plan first, check the numbers myself, and approve it before anything happens."
If you use an agent that can take real actions, such as sending emails or making bookings, the same routine applies, and the checkpoints matter even more. Put one before anything that costs money or can't be undone.
Common mistakes and misunderstandings
Trusting a plan because it looks professional. Numbered steps, bold headings, and precise timings are easy for a language model to produce. They say nothing about whether the plan works. The PlanBench and TravelPlanner results show plans that look right failing on budgets, constraints, and basic possibility. Check the arithmetic and the constraints yourself.
Skipping the plan and asking for the finished job. "Organise my party" in one go gives you a finished-looking answer built on assumptions you never saw. Asking for the plan first costs a minute and shows you those assumptions while they're still easy to change.
Planning everything, even tiny jobs. The opposite mistake. For a one-line email or a quick fact, a plan is pure overhead. Save it for work with several steps, real costs, or things that can't be undone.
Assuming the AI will notice when a task is impossible. In the 2024 Mystery Blocksworld tests, a strong reasoning model usually produced a confident plan for puzzles that had no solution. If your request might be impossible, such as a budget that's too small or dates that can't work, ask directly: "Is this achievable as stated? If not, say so before planning."
Treating the first plan as final. A plan is a starting guess. Good agents, and good people, revise it when a step fails or new information appears. If an agent keeps going after an error as if nothing happened, stop it and ask it to update the plan.
Letting an agent act without checkpoints. An agent with permission to spend money, send messages, or delete files should pause before those steps. Build the checkpoints into your instructions, and use the approval settings your tool offers. The autonomy levels article covers how much freedom to give.
Confusing more thinking with a better plan. Reasoning models with extended thinking do plan better on hard problems, and the research above shows a big jump. They still fail, especially on long plans. More thinking raises the odds; it doesn't replace checking.
Believing old numbers, or new hype. The 0.6 percent TravelPlanner figure is from February 2024, and models have improved a lot since. Equally, a product saying its agent "plans like an expert" is marketing. Judge a planning agent by testing it on your own real tasks.
Related elements
- Agent: planning is what lets an agent take on a whole job instead of a single reply.
- ReAct Loop: the step-by-step think-act-observe pattern that often runs inside a plan.
- Chain of Thought: step-by-step reasoning, the prompting idea that Plan-and-Solve builds on.
- Extended Thinking: models that think longer, which helps most on planning problems.
- Prompt Chaining: a fixed plan you design yourself, one prompt per step.
- Tool Use: the actions a plan is made of when an agent works in the real world.
- Human in the Loop: the checkpoints where a person approves a plan or a step.
- Self-critique: asking the AI to find the flaws in its own plan before you approve it.
- Orchestration: how larger systems split a plan between several AI workers.
- Coding Agents: where plan-first modes are most developed today.