Can AI agents plan multi-step tasks reliably?

It depends on how many steps there are and whether any of them can be undone. Agents plan short, well-defined tasks well. Reliability falls as plans lengthen, because small errors compound and the agent starts trusting its own account of what happened. The fix is not a smarter planner but checkpoints the platform enforces between steps.

Why do longer plans become unreliable?

A plan is a guess about the future made at the start. Each step changes the world a little, and the agent's picture of the world is only as current as its last check. Three things go wrong as steps accumulate.

First, errors compound. A step that is right most of the time, repeated across a long chain, produces a chain that is wrong far more often than any single step. Second, the agent narrates rather than verifies: it records "vendor created" because it sent the request, not because it confirmed the record exists. Third, the goal drifts. Twelve steps in, the agent is optimising for finishing the plan rather than for the outcome the plan was meant to reach.

None of this means agents cannot do multi-step work. It means the reliability has to come from structure around the agent. The mechanics of the loop itself are covered in how an AI agent works, step by step.

When can you rely on an agent's plan?

Plans hold up when:

• The task has a known shape, such as onboarding, reconciling or chasing, and the agent has been certified on real examples of it.
• Each step's result can be checked against a system rather than the agent's memory.
• The steps that cannot be undone come late and are held for a person.
• The plan is short enough that a person can read it in a minute.

Plans are fragile when the task is open-ended, when steps depend on facts that arrive from outside, or when the agent must invent the sequence from nothing each time.

What may the agent do, and what stays with a person?

The first row matters more than it looks. A plan is an output, and for unfamiliar work it should be treated like any other draft. A person reading a five-line plan catches a missing step far more cheaply than a person unwinding the result.

What does a reliable plan look like in practice?

Take supplier onboarding at a manufacturer. The agent receives a request and drafts a plan: collect the supplier's registration documents, check them against the company register, create the supplier record, verify bank details, schedule the first payment.

Under a governed setup, each step passes the platform before it runs. Creating the supplier record is depth 3 work and goes through, after which the agent reads the record back to confirm it exists. Bank details are different. The details arrived in the supplier's own email, so they are external, and the platform holds the step for a person who confirms them by calling a known number. Scheduling the payment is depth 4 and the agent is certified only to depth 3 on payments, so it prepares the schedule and stops. Five steps, two people, no surprises.

If the company register lookup fails, the agent re-plans: retry later, or hold the case. It does not skip the check because the rest of the plan is ready.

What does the platform layer have to do?

• Check every step, not just the first. Authority is evaluated at the moment of each action, so a plan approved at step one does not pre-approve step five.
• Require state checks. The next step does not run until the previous one is confirmed in the system it touched.
• Hold before irreversible steps. See how to design the hold-and-approve step for what the approver should see.
• Limit plan length and re-plans. A task that has re-planned several times is a task for a person.
• Record the plan and every deviation from it, so a reviewer can compare intent with what happened.

Long tasks that pause and resume add their own problems, covered in how AI agents handle long-running tasks. The full design of the governing layer is in the governed agentic operating system guide.