Build an AI Agent with Claude

Catch the recording at 

.

Previously, on Dev Meetups…

At our June Dev Meetup, Charlie Needelman (Principal Product Manager, AI Agents), walked us through a credit hold resolution workflow, and how AI Agents can use Claude.

The Scenario

A customer bought a large MRI system, got flagged for exceeding their credit limit, and it turned out the invoice had incorrectly marked them as taxable — when they're actually a non-profit with a valid exemption certificate. The discrepancy came in through different formats from different systems: a Salesforce record, SAP invoice data, a SharePoint PDF, a ServiceNow support ticket.

If your inputs are consistent, your format is predictable, and your conditions are manageable, a deterministic workflow can handle it. But AI Agents earn their place when things get variable. When a support ticket says "exempt" and an email says "not taxable" and a PDF says something else entirely, a hard-coded lookup fails. An Agent handles it because it can decode the meaning from multiple sources.
 

How the Agent Was Built

The Credit Resolution Agent had four tools to calculate dispute details, call a tax compliance sub-agent, read from shared memory, and calculate credit exposure. A parent Process orchestrated the main agent and a sub-agent that did the heavy lifting on the tax question. The agent read the certificate, compared it against the invoice, confirmed the certificate was valid and unexpired, marked the customer as exempt, recalculated what they actually owed, and handed off to the next step in the bigger process — which eventually surfaces to a human for a final decision.

A few architectural decisions worth noting:

• Sub-agents for complex sub-tasks. The tax compliance check - whether invoices were correctly marked - was handled by a separate child agent with its own model connection and its own context. Cleaner than trying to do everything in one agent, and it ensures the tax logic is reusable elsewhere.
• Shared memory as a data layer. After the sub-agent made its decision, it wrote the results to an Azure storage table. The parent agent then read from that table to proceed. This was an experiment with a pattern that's more useful in longer, multi-step processes where you don't want to thread data all the way through.
• Scope your inputs. In the demo, Charlie sent in one large JSON per input variable, and immediately flagged it as the wrong call. The more context an agent gets that it doesn't need, the more likely it is to hallucinate. For anything you're building for real: break your inputs into individual variables, map only what the agent actually uses, and cut the rest.
• Build a fallback path in two places. In the action plan, and in the tool description. And if you don't want to escalate to a human every time, route it to a queue, fail gracefully, retry a few times first.

Q&A Highlights

How strictly does the AI Agent follow negative constraints — and does Claude try to work around them?

The first guardrail is just to limit the tools, as AI Agents can only pick from the tools you give them. The Agents in Automation Anywhere are capped at 10 tools for exactly this reason. so keeping that list tight is your first line of defence.
Beyond that, make your constraints specific: "Escalate if you get an unexpected response" is weaker than "if the tax compliance agent does not respond as expected, escalate immediately." The latter leaves less room for interpretation.

How can I use Claude as the model for my AI Agent?

Claude Sonnet 4.5 is available as an Automation Anywhere-provided model connection out of the box, which is what Charlie used for this demo. More models are being added more regularly as AA revamps its model lifecycle process — the goal is to get new model versions out closer to when they're released publicly.

How can I use AI Skills in AI Agents?

Currently, if you want to use an AI skill in an agent, it has to be wrapped in an API Task/Taskbot first. That's changing in the upcoming releases, as we have direct skill-as-tool support for AI Agents.

Pro Tip

If the AI Agent isn't running a tool you need it to run, make sure the final output/next tool in the step requires a parameter that can only be obtained as a result of running that tool. Prompting agents is different to prompting a general-purpose LLM. The action plan, the tool descriptions, the output variable names and their descriptions — all of it is context the agent uses to decide what to do next. If any of it is vague or contradictory, you'll see it in the agent's behaviour.

🔮 What's Next

The next dev meetup is on August 27th. Winners of our Pathfinder Hackathon will showcase their demos.

Sign up here to watch them: 

Want to present a use case, a half-formed idea, or anything in between? Reach out at 📧 community@automationanywhere.com