Agent step

The Agent step lets a model reach a goal in several steps. You describe the goal in plain language and name the tools it may use. At each step the model reads the situation, chooses a tool, reads the result and goes on until the goal is reached or the step limit is.

Flow or agent

	Flow	Agent step
Logic	You decide every step once.	The model works it out on each run.
Same input	The same steps, always.	The model may act differently.
Use it when	You know the process.	The work varies and cannot be drawn as steps.

If you can write the rule in one sentence without saying “depends”, it is a condition in a flow and not a job for an agent.

Add the step

Select Add step, open Collabase AI and choose Agent.

Write the goal

Say what the agent should achieve and what a good answer looks like. Insert data from earlier steps where it matters. The goal reaches the model as data to work on, not as an instruction it must obey blindly.

Add tools

Add the tools the agent may use. An agent can only do what its tools allow.

Choose approvers

If any tool asks a person first, choose the approvers. They can be people and groups.

Test the step

Run Test step. The agent runs for real, so tools that change something do change it.

Inputs

Input	Description
Goal	What the agent should achieve.
Tools	What the agent may call.
Fixed inputs	JSON with values the agent must not change, such as the project of a new work item. For example `{"collabase-tasks.create_task": {"projectId": "..."}}`. The key is the plug and the action.
Approvers	People and groups who approve tool calls that ask first.
Most steps	How many steps the agent may take. The default is 20, the most is 50.
Model	Model of the instance or Provider of this workspace. See AI provider.

What the step returns

Field	Content
`text`	The agent’s final answer, a short summary of what it did and found.
`steps`	How many steps it took.
`tokens`	The tokens it used.
`truncated`	`true` if the agent stopped at the step limit before it was done.
`calls`	One entry per tool call with the tool, whether it worked and, for tools that asked first, whether it was approved.

The text varies from run to run. For branching, ask in the goal for a short, fixed answer, or let the agent call a flow tool whose steps do the work.

Limits of an agent

An agent has no memory between runs.
It never invents results. If a tool did not return something, it says it does not know.
If a tool is refused or fails, it reports that and does not look for a way around.
Tool results are long text, so Collabase cuts a result to 8,000 characters before the model reads it.

Give an agent only the tools it needs, and ask for approval for every tool that changes something or sends something. Text from outside, such as a search hit or the data a flow started with, may try to give the model orders. The step hands it over as data, but approval is the safeguard you control.

Approvals while the agent works

When the model chooses a tool that asks first, the run pauses. The approvers get an entry in their inbox with the tool and the exact input. They approve or reject within seven days. Only the input they saw runs, and the agent goes on with the result or with the refusal.

Track the cost

Agent steps and the tokens they use appear in the workspace analytics. A workspace that uses its own provider can set a monthly token budget that stops AI steps when it is used up.

Overview

Build flows

Run and publish

Agents and AI

Share and reuse

Workspaces and plugs

Reference

Build a plug

Plug triggers

Plug runtime

Ship a plug

Flow or agent

Add the step

Inputs

What the step returns

Limits of an agent

Approvals while the agent works

Track the cost

​Flow or agent

​Add the step

​Inputs

​What the step returns

​Limits of an agent

​Approvals while the agent works

​Track the cost

​Related

Flow or agent

Add the step

Inputs

What the step returns

Limits of an agent

Approvals while the agent works

Track the cost

Related