Multi-step AI agent tasks: plans, helpers and handing parts to other AI employees
Some jobs are a list of questions, not one. How an agent plans a big task, works through it, delegates parts, reports back, and when a fixed flow is the better choice.
Part of the guide AI employees
Multi-step AI agent tasks are jobs where the agent cannot answer in one go: it has to plan what to find out, work through the steps one at a time, use different tools along the way, keep notes, and finish with a report. In xMap Flows you get this by setting the AI agent's How it works to Work through a big task, giving it tools and, if you like, other AI employees it can hand work to through Give work to. Afterwards the run shows the plan as a checklist, so you can see what it did at each step.
This is for analysts, consultants and managers who keep doing the same open-ended research by hand: comparing areas, checking a list of sites, preparing a weekly brief. The glossary entry on the AI agent covers the basic idea; here we go deeper into planning and delegation.
A plan, then steps
In this mode the agent brings its own starting item, so a Schedule can come straight before it. It reads the task in its instructions, writes itself a to-do list, works the steps in order, and may hand a self-contained step to a helper when that is simpler. Between steps it keeps notes, so step six knows what step two found. It ends with a report in {{report}}, which an Email or Make a report block after it can send.
- Most steps (3 to 30) is the longest to-do list it may work through, including steps it adds as it learns.
- A long task carries on across the run's parts, up to about an hour. The run shows the step it is on, such as Step 4 of 9.
- It stops at Most steps or when the run's credits run out, and still reports what it has, saying why it stopped.
- A test run stops after three steps and says so, so you can check the plan's shape cheaply.
Handing parts to other AI employees
Give work to lists up to eight of your other AI employees. Each becomes a tool the agent can call, such as Ask the Opening watcher: it runs that employee's flow with what the agent gives it and brings back what it found. This works whether the agent answers messages or works through a big task, and it lets a manager agent reuse employees you already trust rather than rebuilding their logic.
Sending stays controlled. An employee that sends something never sends while the agent is thinking. In a test run nothing is sent; in a real run that sends, each employee's sending happens once, after the agent has finished. Each employee's run uses its own credits within what the manager's run has left, and appears in its own run history. You cannot give work to an employee that is archived or paused, private to someone else, or contains a Wait or an Approval.
A worked example (made up)
A healthcare consultancy in Sao Paulo prepares a Monday brief for a client choosing where to open its next physiotherapy clinic among five neighbourhoods. Before, an analyst spent half a day on it. The AI employee:
- Schedule: every Monday at 06:00.
- AI agent, Work through a big task. Instructions: For each of Pinheiros, Moema, Tatuape, Santana and Butanta, count physiotherapy clinics and gyms, note anything that opened recently, and summarise local news about health services. End with a ranking and one paragraph per neighbourhood. Most steps: 18.
- Tool: Nearby places, its point written as {{agent: the neighbourhood to look at}}, types physiotherapy and gym.
- Tool: News, reading a few local health and business feeds.
- Give work to: the firm's existing Opening watcher for Sao Paulo, which already tracks new clinics.
- Make a report: a PDF with {{report}}, emailed to the client team.
The agent's plan comes out as ten steps: one count per neighbourhood, one call to the Opening watcher, two news reads, a comparison and the write-up. One count returns little, so it adds a step to search a wider radius. The PDF lands at 07:00 with a ranking and a paragraph each, and the analyst spends twenty minutes checking it instead of half a day producing it.
Reading the trace
Open the run and press See how it answered on the agent's line. The plan appears as a checklist: done, nothing found, or not reached. Open a step to see each tool call, what it was given and what came back, a helper's work folded inside, the notes, and the time and credits of each step. Then comes the report and why it stopped, such as Stopped: Most steps. Working notes are shown only to the flow's owner and the organization's admins.
When a plain flow is better
If the job is the same sequence every time (search, filter, score, email), build it as an ordinary flow. It is cheaper, faster and easier to check, because every run does exactly the same steps. A big task earns its place when the steps depend on what earlier steps found, or when the questions themselves change week to week.
Mistakes to avoid
- Vague instructions such as research the market. Name the places, the measures and the shape of the report.
- Most steps set to 30 for a task that needs ten; it caps cost, so set it close to what the task needs.
- Giving it every employee you have. Pick the two or three whose findings it really needs.
- Deploying without a test run: the first three steps show whether the plan makes sense.
Ready-made examples include the research assistant over live location data and companies hiring transformation leaders, briefed by an AI agent. The AI employees guide shows how manager agents sit among the rest of your team.
Questions
How long can a big task run?
It carries on across the run's parts for up to about an hour, and stops sooner at Most steps or when the run's credits run out.
Does it still report if it stops early?
Yes. It always writes a report with what it has, and the trace says why it stopped.
Can the agent make another AI employee send emails?
Only in a real run that sends, and only once, after the agent has finished. In a test run nothing is sent.
Which employees can it give work to?
Up to eight of your other AI employees, as long as they are not archived, paused, private to someone else, or contain a Wait or an Approval.
Templates that do this
Read next
More on ai employees and automation · Every article · Templates