Lesson 3 of 9 · 9 min read
ReAct vs plan-first
Some agents think one step at a time and adapt as they go. Others write the whole plan first and then work through it. Knowing which pattern fits a task saves time, money and mistakes.
Two ways to get from goal to done
Give two capable people the same task and you may see two styles. One starts immediately, checks what they find, and adjusts as they go. The other spends ten minutes writing a list, then works through it without stopping. Neither is always right. Agents have the same two styles, and they have names.
ReAct: think, act, look, repeat
ReAct (short for reasoning and acting) interleaves thinking with doing. Each turn of the loop has three parts:
- Thought: "I need last month's churn figure. The billing system should have it."
- Action: call the billing report tool for August.
- Observation: "Churn was 4.1%, up from 2.3%. Most cancellations are on the Starter plan."
Then a new thought, based on what it just learned: "Why Starter? Check support tickets from Starter customers in August." The next action depends on the last observation. That is ReAct's strength: it adapts. It follows leads, recovers from surprises and handles tasks where nobody knows the steps in advance.
The cost is also in the loop. Every step sends the whole growing history back to the model, so long runs become expensive and slow. It can wander, chasing interesting but irrelevant leads, and it is harder to predict what it will do, which makes it harder to review before it starts.
Plan-first: write the plan, then work it
A plan-first agent separates thinking from doing. The best-known pattern, ReWOO (reasoning without observation), splits the work into three roles:
- Planner: reads the goal once and writes the full plan, with placeholders for results it does not have yet. "Step 1: get the receipts (#E1). Step 2: extract vendor, date and amount from #E1 (#E2). Step 3: categorise #E2 (#E3)."
- Workers: carry out each step's tool call and fill in the placeholders.
- Solver: takes the plan and all the evidence and writes the final answer.
Because the model is not re-reading a growing transcript on every step, plan-first uses far fewer tokens. The ReWOO authors reported roughly five times fewer tokens than ReAct on a multi-step question-answering benchmark, with similar or better accuracy. Plans are also predictable and reviewable: a person can read the plan before anything happens.
The weakness is the mirror image of ReAct's strength. If step 2 returns a surprise, the plan does not know. Plan-first is weak on exploratory tasks, where the right next step depends on what you just found.
How to choose
Ask two questions: do we know the steps in advance? and what does a mistake cost?
| The task | Better fit | Why |
|---|---|---|
| Fixed, repeated steps (file expenses, onboard a supplier, weekly report) | Plan-first | Steps are known; cost matters at volume; the plan can be reviewed once |
| Open questions (why did churn jump, debug a failing job) | ReAct | The next step depends on what you just found |
| Several changes to real systems | Plan-first with review | A person should see the whole set of changes before any happen |
| Mostly known, with surprises | Hybrid: plan, then replan | Start from a plan; when a step returns something unexpected, revise the rest |
The hybrid is what most production agents do in practice. They decompose the goal into steps, sequence them, and replan when reality disagrees with the plan. Planning is not a one-time act; it is a habit of checking the plan against what happened.
Worked example: same agent, two requests
Request A: "File my three receipts from the Sydney trip as expenses." The steps never change: read each receipt, extract vendor, date and amount, pick the category, create three expense lines, submit. A plan-first agent writes that plan in one pass, you glance at it ("three lines, travel and meals, total $412"), approve, and it runs. If this happens for fifty people a month, the token saving is real money.
Request B: "Our Starter plan cancellations doubled in August. Find out why." Nobody knows the steps. Maybe it is pricing, maybe a bug, maybe a competitor. A ReAct agent checks billing, notices most cancellations came after a failed payment, pivots to the payment provider's logs, and finds a card-update page that broke on 3 August. A fixed plan written at the start would never have looked there.
The lesson: pick the pattern per task, not per agent. Many agents can do both, and the best ones choose.
Plan review: a human checkpoint before the work
Plan-first has a governance bonus. Because the plan exists before anything happens, a person can review it: remove a step, change a target, reorder the work, or stop it altogether. That is cheaper than approving ten separate actions one by one, and it catches a bad approach before it has done anything. It does not replace approval of individual actions that change data; it adds an earlier, wider check.
Renly supports both styles. For moderate and complex requests, the agent is guided to plan its steps first. An agent can also put its plan up for review, and when an agent has Plan review before execution switched on, it pauses and waits for you before carrying out the plan.
Review a plan, then compare with a one-step request
About 10 minutes. You need an agent you can edit. This exercise uses Renly's own to-do tasks, so no connector is needed.
- Open the agent in the builder. On Identity & Personality, under Execution settings, switch on Plan review before execution. Save.
- Choose Chat and give it a multi-step request that changes things, for example: "Plan next week's product launch: create a to-do task for the press release, one for the customer email and one for the internal briefing, each with a due date, then summarise the plan."
- The Plan panel opens with the goal and the steps. Edit the Goal or a step's Description, remove a step you do not want, and move one Up or Down.
- Choose Run edited plan (it reads Approve & run if you changed nothing). Depending on the agent's autonomy, each task it creates may still ask for approval on its own card.
- Now ask a one-step question: "What is on my to-do list this week?" Notice there is no plan to review. A single read does not need one.
Choose Cancel on the plan panel at any time to stop before anything runs. Switch plan review off again if you only wanted it for this exercise.
Key takeaways
- ReAct loops through thought, action and observation. It adapts well but uses many tokens and can wander.
- Plan-first (ReWOO style) plans once, then works the plan. It is cheaper and reviewable but weak on exploratory tasks.
- For fixed-step work where cost matters, choose plan-first. For open questions, choose ReAct.
- Most real agents plan, then replan when a step returns a surprise.
- A plan is a checkpoint: reviewing it catches a bad approach before any action happens.