What happens when an AI Agent needs to read a lengthy contract, check information in external systems, and apply a calculation rule? This work brings together different tasks. Execution needs capabilities suited to each one, with monitoring and human involvement where needed.
The improvements presented at Pipefy Updates cover that path from validation before publishing to investigating what was executed. You will see how testing, connections, calculations, document analysis, and human approval can work together as agents take part in your processes.
Validate behavior before putting it into operation
Predictability starts before execution. Behavior testing uses a card from the pipe as a reference to simulate what the agent would do, without changing production data. IT and process owners can review the result, refine the instructions, and test again before publishing.
Improve instructions supports that review. It considers the agent's fields and actions to suggest changes, while the person responsible chooses which suggestions to accept, reject, or edit. Agent guidance remains under the control of the people who understand the process.
Testing also covers automations that move cards, create cards, and create connections. Simulations check the actual pipe configuration without changing production data and identify failure reasons, such as a missing required field or an incorrect phase.
The cycle is to write, test, refine, and test again until the result meets the process requirements. Model choice is also part of preparation: a simple task can be evaluated with Pipefy AI Lite, while a more complex activity may require different capabilities. Models are selected by behavior.
Explore behavior testing, in Portuguese, Improve instructions, in Portuguese, and testing card movement automations, in Portuguese.
Connect the agent to information in other systems
MCP Client lets AI Agents call external tools directly within their behaviors. In a sales process, for example, the agent can look up a lead in Salesforce or HubSpot, bring that information into Pipefy, and continue working with that context.
The managed connector catalog supports authentication through OAuth or an API key. Configuration also lets teams control which tools each agent can use. Access can therefore be scoped to the work the agent needs to perform.
Explore MCP Client, in Portuguese, and connectors for AI Agents, in Portuguese.
Run calculations and business rules with code
When a decision depends on numbers, how the result is produced matters. Calculations & Analysis lets an agent write and execute small programs in a secure environment.
This capability supports calculations, business rules described in natural language, and extracting data from PDFs and images into structured fields. Numerical operations are performed through code, with logic that can be tested against known examples.
A commercial rule might apply a 10% discount when sales reach a quantity defined by the company. The agent needs to check whether the criterion has been met and apply the corresponding calculation. The team can test eligible and ineligible cases to check both the result and the interpretation of the rule.
Learn about Calculations & Analysis, in Portuguese.
Look deeper into lengthy documents
Some tasks require finding information spread across a document and connecting passages before taking action. Max Effort supports deeper reasoning: the agent plans, searches for information, retrieves specific pages, and revises its plan during execution.
Available with Pipefy AI Pro, the capability is activated by behavior. It supports working with documents longer than 30 pages, gathering the information needed for the task, and populating fields according to the instructions. The analysis can connect to subsequent behavior actions, such as sending results through Slack.
Explore Max Effort for complex workflows, in Portuguese.
Keep people involved and follow the execution
In a reimbursement process, a company can define an amount below which a request proceeds automatically and require human approval above that limit. The agent sends a task by email for the responsible person to approve or reject the proposed action. It can also request validation when its instructions require confirmation about the applicable policy.
Logs show the status, time, and steps of each execution, including information consulted, fields updated, and AI Credits consumed by each action. This helps teams review the execution path and resource usage. Dynamic translation of tracing nodes follows the Pipefy profile language, with an English fallback.
Validation and tracing complement each other: one establishes human participation in the decision; the other helps explain how execution reached that point.
Explore Human Validation, in Portuguese, and tracing node translation, in Portuguese.
Use Pipefy tools from other environments
For technical teams, MCP Server expands the ways to access Pipefy. The toolkit enables its tools to be used from environments such as Cursor, Claude Code, Claude Desktop, and Codex, connecting work in those environments to company processes.
With MCP Server, your team can ask a compatible assistant, such as Claude, which cards are pending in a pipe without opening the Pipefy interface. This approach, known as headless access, connects work started in other tools to company processes, subject to applicable access controls.
Learn about MCP Server and the toolkit, in Portuguese.
Expand execution with context and monitoring
Testing before publishing, connecting the necessary sources, choosing suitable capabilities, and defining human involvement shape how agents work. Execution records help teams follow that work and guide adjustments.
Keep exploring Pipefy Updates
Return to the event recap or explore the other journeys:
