Build process agents

A process agent runs your process the same way every time. You set the steps and the rules between them, AI agents do the reasoning and evaluation inside the steps you choose, and every run records the route it took. This article explains what a process agent is, how to build one with Omni, how it stays on its route and what it can't do yet.

What is a process agent?

An autonomous agent picks its own next step, so you can't promise it will take the same steps on every case. A process agent follows the steps you set, in your order, and each decision picks from a few set answers, like approve or reject. The AI works inside the steps, reading documents, comparing records and making a call, and your rules read its answer to decide where the work goes next. The same case always takes the same route, and the run record shows why.

Use a process agent when work must run the same way on every case and leave a record: approvals, claims, payments, onboarding and compliance checks. Use an autonomous agent when the path isn't known up front, and an agent team when the work spans several skills. See Build and deploy agents.

Key capabilities

• Mixed steps: a step can be a few lines of code, one model call, a tool, a full agent with its own tools, or an inspector.
• Rules you set: a rule reads what earlier steps found and picks the route, and every rule has a default route for the case nobody expected.
• Checked before it runs: aixplain refuses a process with a rule that has no default route, a step that can't be reached or a loop you didn't allow.
• Loops where you allow them: work can go back to an earlier step, with a limit on how many times it goes around.
• Retries and time limits: a step that fails tries again, and a slow one times out.
• Inspectors and budgets: inspectors check an AI step's answer before a rule acts on it, and each run has a limit on cost, time and steps.
• A record of every run: each step, the route each rule picked and what the run cost.

Build a process agent with Omni

Describe the outcome and the rules, and Omni builds the process agent for you. Start from the process you already have written down in a policy, a process tool or a script.

1. Open aixplain and select Omni.
2. Describe the outcome you want and the rules the work must follow, such as approval limits, policy checks and who handles the exceptions.
3. Answer Omni's questions as it builds the process agent. It turns each rule into an instruction for an AI step or an inspector that checks it, depending on how strictly you need the rule enforced.
4. Review the rule each AI step follows, and tell Omni what to change. Each rule comes from your description, so you change it by changing the description.
5. Test the process agent on several cases, including ones that don't fit the rules, and refine the rules until every case takes the route you expect.

Example requests

• "Build a process agent for supplier invoices. Match each invoice to its order, send price mismatches to a step that checks the supplier's latest price list, and hand anything over $10,000 to the finance lead."
• "Turn our claims policy into a process agent. Claims over $5,000 always get a fraud review, and high-risk claims go to an adjuster."
• "Run these ten past claims through the process and show me any that took a different route than the adjuster did."

Bring a person in

A process agent doesn't pause to wait for a person. When a case needs one, such as an invoice over the approval limit, the process hands it to that person with what they need, like the match report, and the run ends there.

Call a process agent from another agent

An autonomous agent or an agent team can hand work to a process agent. A support agent can talk a customer through a problem in its own way, then pass the refund to your refunds process, which handles it the same way it handles every refund.

Build it in code

Developers can build the same process agent in Python with the Agents SDK, step by step, and call it from their own apps.

Run and deploy it

A process agent is an agent like any other: it has an ID, runs on cloud or on your own servers, and can be called from the API, a schedule or a trigger. See Deploy an agent and Run an agent on a schedule or a trigger.

Pricing and usage

A process agent costs what its steps use. Each model, tool and agent step draws credits from your workspace, while code steps and rules don't call a model. For plans and rates, see Your account, plans and billing and Pricing.

Limitations

• There's no visual editor for process agents yet. You describe one to Omni, or a developer writes it in Python.
• A step can't wait for a person to approve. The process hands the case over and the run ends.
• A process agent can't be used as a step inside another process agent yet.

Frequently asked questions

Get help

• Contact us: write to care@aixplain.com.
• See what changed: read the release notes.