👤 For process owners and pipe admins
🔐 AI Agents are available on all plans, subject to company enablement, permissions, and credits
🎯 For people who want clear responsibilities and a way to check each AI task's output
An expense reimbursement request arrives with a receipt and an explanation. Checking the information, comparing the expense against company policy, and authorizing the request are different jobs. When those outputs are unclear, the person following the request has to retrace the work to understand what is complete and what still needs review.
By the end of this article, you will be able to decide when to keep a task in one behavior, add behaviors to the same AI Agent, or create separate agents. You will also know how to define each part's output and the conditions that allow the next part to begin.
What you will learn:
Identify the outputs within the work
Before deciding how many agents to create, describe the result you need at each point. “Handle the reimbursement” is broad. “Identify the information on the receipt and flag anything missing” gives you an output you can check.
Ask three questions: what information comes in, what work needs to happen, and what result shows it is complete? A lookup might provide a reference; an analysis might deliver a recommendation with a rationale; an update might record the result in a card field.
Not every part requires interpretation. When a stable, objective condition determines the next action, consider a rule or a complementary automation. Use AI analysis where the task requires interpreting information in the context of the process.
If you still need to define the task, see how to choose a task and its inputs, in Portuguese. Then describe the output you expect to check.
Choose the right approach. The Panorama da IA course, in Portuguese compares Rules, AI Automation, and AI Agents, and covers control, data, and costs. Use those criteria to decide where AI fits your routine.
Choose another behavior or another agent
In Pipefy, an agent has a function and reference knowledge. Behaviors define situations in which it acts, with their own triggers and instructions. Think of the agent as a role in the process and a behavior as one way that role acts.
For example, a support triage agent could classify a request when it arrives and reassess it after the requester adds information. If both activities serve the same role, they can be organized as behaviors within the same agent.
Use this comparison to decide how to divide the work.
| What you need to organize | Division to consider | What to check |
|---|---|---|
| One coherent output with clear inputs and results | One behavior | The instructions serve one purpose and fit the feature's limits |
| The same role acting at different points or events | Behaviors within the same agent | Each behavior has its own trigger and expected output |
| Responsibilities with different references or acceptance criteria | Separate agents | Each output can be explained, tested, and maintained separately |
| An action determined by an objective condition | A rule or complementary automation | The owner of the action and the condition that allows it are clear |
Separate agents make sense when they help organize responsibilities. Document checks may use different criteria from a policy assessment. Creating an agent for every sentence in an instruction adds more parts to maintain without necessarily making responsibilities clearer.
Different models can stay within the same agent. Models are selected per behavior. Needing different levels of analysis does not, by itself, require another agent. See the reference on models per behavior.
Treat limits as boundaries, not targets. The documentation supports up to five behaviors per agent and up to three core actions per behavior. If the work goes beyond that scope, review the division and its dependencies. Refer to creating an AI Agent.
Define how work passes to the next stage
Every transition needs an agreement. If an analysis depends on checked information, define where that information will be recorded and which event allows the analysis to begin. Naming the parts “checks” and “analysis” does not establish that dependency.
For each part, record:
- Required inputs: information and references that must be available.
- Verifiable output: the expected result and where it is recorded in the process.
- Continuation: the event and condition that allow the next part to start.
- Exceptions: what to do when information is missing or conflicting.
- Ownership: who checks the output and resolves outstanding issues.
If two parts could update the same result or move the same card, decide which part owns that action. This makes it possible to identify the source of a change and avoids relying on different interpretations of who should perform it.
When human review is required, continuation should depend on the recorded decision. The documentation shows a subsequent behavior triggered by an update to the field completed by the reviewer. Also define which decision allows progress and which requires changes. See chaining behaviors after human validation.
Represent the division in the process. Do not rely on the order of behaviors in a list to establish a sequence, or assume that one agent receives another's output without configured inputs and events. Separate agents also do not replace access and approval decisions.
Apply the decision to an expense reimbursement request
Imagine a team receiving receipts and explanations for expenses. This is a hypothetical example: adapt the names and records below to your company's policy and process.
The first output identifies information present on the receipt and flags anything missing. The second compares the available information against policy and provides a recommendation with a rationale. Authorization stays with the person designated by the company.
If the same analysis support role performs an initial check and reassesses the request after additional information arrives, consider behaviors in the same agent. Each point needs a defined event and an output that can be checked.
If document checks and policy assessment have distinct references, acceptance criteria, and maintenance needs, consider separate agents. One can record the information found and any outstanding issues; the other can use the recorded information to prepare its recommendation. The division is complete only when the process defines how the second part receives the information and when it may act.
To represent the transition, you might plan records such as “Checked information,” “Outstanding issues,” and “Review decision.” These are illustrative names, not required product fields. If the amount on the receipt is unreadable, the expected output is an issue for someone to check, without estimating an amount.
Before continuing, agree on who requests additional information, how corrected information enters the analysis, and who decides when information conflicts. Preparing a recommendation and authorizing a reimbursement remain distinct responsibilities.
Use the division to find the output that needs review. If the receipt information was identified correctly but the recommendation does not match policy, you know which result to compare with its references and instructions. More agents help only when this distinction is clear.
Test each output and the complete process
Start with cases where you know the expected result. Compare each part's output against the source information, then follow the complete process, including requests for additional information and human decisions.
| Test case | What to check |
|---|---|
| Complete request | Each part receives the required information and produces the expected output |
| Incomplete receipt | Missing information is recorded as an issue rather than treated as known data |
| Conflict with policy | The recommendation explains the conflict and goes to the designated review |
| Additional information | The new analysis uses updated information and does not repeat a final action unnecessarily |
| Approval and rejection in review | Each decision follows the agreed path |
AI Agent Tracing and Logs help you investigate a run. Compare the records with the expected output: completing a run does not, by itself, prove that the analysis is correct.
Repeat these checks after changing instructions, references, or dependencies between stages. The division is useful when you can explain who delivers each result, what allows work to continue, and where a person participates.
Before moving on, confirm that you understand:
☐ I can distinguish an agent's role from the work performed by its behaviors.
☐ I can explain why to keep responsibilities together or separate them in my routine.
☐ I have defined each part's inputs, output, continuation event, and exception handling.
☐ I know who checks the outputs and participates in review.
☐ I have cases to test each part and the complete process.
Ready to check this division in your own process? → Continue with testing and monitoring processes with AI, in Portuguese.
