AI Agent Action Type

The Fully Autonomous AI Agent action type hands a task to an AI agent that works on its own. Unlike a simple AI Prompt, which sends one prompt and uses one answer, the agent works in steps: it looks up information, runs code, checks the results, and keeps going until the job is done (or it determines the job cannot be done).

While working, the agent can:

Look up your process templates, running processes, tasks, and process table records
Update process fields and start new processes
Write and run JavaScript code in a secure sandbox - including calling external APIs and reading web pages
Find API keys you have stored as environment variables (Administration > Integrations > Environment Variables) and use them in its code
Send a question to a specific person and pause until they respond, then pick up where it left off
Choose the task response that best matches its findings and complete the task itself

You can also have this set up for you: describe the work to the AI Strategist - “have AI draft the reply on the triage step” - and it writes the agent’s assignment, picks the field to store the answer in, shows you the whole setup, and puts it live once you agree. The steps below are for building one yourself.

To set up an AI Agent action:

On the desired template, click the Automated Actions icon in the template designer menu
Give the action a name and select Fully Autonomous AI Agent as your Action Type
Select a trigger and add any conditions
In Action setup, complete the settings below

AI Agent settings

AI Role Description - An optional description of the role the agent should play (for example, “You are a thorough AI financial analyst”)
Natural Language Prompt - The assignment for the agent. The more detailed and specific the prompt, the better and more reliable the results. Use the Insert Field Token button to include values from your process
AI Model - Choose the level of intelligence needed
Store Result in Field - Optionally select the process field in which to store the agent’s final answer
Task to Mark Completed - Optionally choose a task for the agent to complete. The agent reviews the task’s available responses and chooses the appropriate one on its own
Response if the Agent Can’t Finish - An optional fallback response, used only if the agent errors out or never selects a response (for example, a “Needs Review” branch). Leave blank to do nothing

Granting external tools

The Action setup step also includes a tool-grant section for your MCP connections (set up under Administration > Integrations > MCP Connections). The agent can only use the external tools you explicitly grant to it here - everything else stays out of reach.

Good to know:

The agent works under its own name, AI Agent, not as the person whose task set it off. Field updates, notes in the audit log, and anything else it does are recorded as the AI Agent, so nobody is credited with work they did not do. It can see and work with everything in your account, except a process whose Visibility List limits it to specific people or groups
Some decisions always need a person, so the agent never takes them: it does not answer the AI Strategist’s questions or give it guidance, approve or dismiss the Strategist’s recommendations, grant the Strategist permission to act on its own, change the Strategist’s focus areas, set up automations, change a process’s status, or edit your company profile. Those happen in a conversation between a person and the AI Strategist
The agent also never leaves behind something other work depends on. It can call the JavaScript functions in your library and run code of its own in the sandbox, but it cannot create or change a function in the library, and it cannot add to what the AI Strategist knows about your company. Because an agent often reads outside content, such as web pages, this keeps anything in that content from lingering after the agent’s work is done
Dates and times the agent works with are in the time zone most of your team uses. The current user field tokens refer to the AI Agent itself, so to mention the person involved, use the tokens for the assigned user or the process initiator instead
The agent works through a limited number of steps per run, and each JavaScript run has its own time limit, so a runaway job cannot loop forever
When the agent pauses to ask somebody a question, it does not wait forever. If nobody has answered after a week, it stops waiting and the task takes the Response on Failure you chose for the action, so the process carries on down that branch instead of sitting still. A job the agent hands to an AI Employee is given a day to report back before the same thing happens. This is worth planning for: point that failure branch at a step that asks a person to pick the work up, so an unanswered question turns into somebody’s task rather than a dead end
When the process is running in Test Mode, every step the agent takes - its reasoning, the tools it uses, and what those tools return - is written to the Instance Audit Log, so you can see exactly how it reached its answer. This is the place to look when an agent does something you did not expect
AI Agent actions count toward your account’s AI usage. If the account’s AI allowance has been used up, the action will not run until the allowance resets