March 2026 Dev Meetup Recap: Choose Your Own Workflow with AI Agent Studio

Catch the recording at https://youtu.be/39LMEkEpQLw?si=qeRM1ZHkGO2af9Hq

Previously, on Dev Meetups...

We did things differently this time around. By now, most of you are familiar with AI Agent Studio and what it can do. So this time, we handed the agent a full toolkit and no instructions on how to use it, let the community vote on the design decisions, and watched it figure out the rest.

The Setup

The session was built around AI Agent Studio and a fictional retail dataset: Max's Bargain Homewares, a fully relational customer support application with order history, a product catalog, complaint records, and a work queue. Matt had pre-built a set of API task tools wired to this backend. The community helped us decide what kind of agent to build with them.

Before getting to the polls, we walked through something worth sitting with: the difference between automating with a legacy mindset versus an agentic one. The legacy approach scopes by team — build the customer investigations process, then product analysis, then testing. Each project is its own thing. The agentic approach does the opposite: look across the entire process map first, identify the shared patterns, and build tools abstract enough to serve all of them. The same getNextWorkItem tool that serves a customer investigations analyst serves a product review analyst. You build it once, and the agent figures out how to use it.

If you'd built those tools specific to any one horizontal use case, you'd have to rebuild them for every new prompt you sent the agent.

Choose Your Workflow 🗳️ [00:17:07]

Five polls, five design decisions. Here's what we designed together:

• What question does this agent exist to answer? Which products are costing us the most?
• Which team is running this agent? Finance
• How does the agent surface its findings? Flag for human review via AnswerForm
• What signals does it use? Volume and history — complaint count and order patterns
• What happens when data isn't there? Retry once before continuing

Behold: a finance-team agent for Max's Bargain Homewares, scoring products by complaint rate relative to units sold, flagging findings via AnswerForm for a human reviewer, with retry logic on tool failure.

The Build — and the Break [00:25:00]

With the community, we built the getNextWorkItem API task live and then dropped the community-generated prompt into AI Agent Studio. Skip to 00:25:00 to find out how that turned out.

Spoiler alert: It broke immediately.

🌟 Matt's Side Experiments at Max' Bargain Homewares [00:37:18]

While the main agent was running, @matt.stewart had been making moves in silence, throwing random queries at the same toolset. No action plans reviewed, no guardrails: just a goal and the agent to make its own connections.

The most interesting one: "For every historical complaint made by an abusive customer, how much did we have to refund because they were recalled products?" The agent figured out how to involve the human in the loop and prompted Matt to enter missing information that it required to determine and execute its next steps. Here’s how that went 00:37:18

Key Takeaways

✅ Build tools for reuse, not for workflows — Abstract, single-responsibility tools serve more agents than use-case-specific ones ever will.
✅ The execution log is your debugging interface — When the agent calls a tool wrong, the log shows exactly what inputs it passed (or didn't). Start there.
✅ Prompt generation gets you 80–90% of the way — The Generate button in AI Agent Studio reads your tool descriptions and variable descriptions to build the action plan. Write those well.
✅ Human-in-the-loop by design — The community voted for AnswerForm over auto-reporting by a clear margin. Flag findings, don't write them directly to the system.
✅ Agents ask when they don't know — Given a well-designed toolset, the agent will surface gaps rather than hallucinate past them.

🔮 What's Next

• The next Dev Meetup is on May 28th — a power workflow built from everything released in the latest product drop. Register via the Pathfinder events page.
• Also: AI Automation Engineer Certification with agents and tool design as core components is coming within the next two months. Hit Join the Waitlist on the Automation Anywhere University page to be first in line.

Want to present a demo at a future Dev Meetup? Reach out to us at 📧 community@automationanywhere.com