How Automation Developers Are Building Goal-Based AI Agents?

There’s a noticeable shift happening in automation.

For years, teams optimised workflows. Now they’re starting to design goal-driven systems. These systems are powered by AI agents that interpret context, navigate ambiguity, and coordinate decisions across tools.

We launched the Agentic Bounty Challenge to unleash the creative problem solving of Pathfinder community members. The rules of the challenge were simple: solve a real-world business problem by designing and building goal-based AI agents. The AI agent needed to have at least three tools and interact with two or more applications (beyond LLMs).

Across all the submissions, we were highly impressed by the outcomes-oriented agent design, and the storytelling that went along with it. This piece talks about the three builds that stood out.

👉 Watch the full episode to meet the winners of the Agentic Bounty Challenge on YouTube or listen on your favorite audio platform.

🥉 Third Place: Jenna Saunders – Turning Product Reviews Into Insight

Jenna approached product reviews as a design problem rather than a data problem. Instead of treating sentiment analysis as a reporting feature, she structured it as an end-to-end agent workflow.

The agent reads review text, assigns sentiment scores, generates structured CSV outputs, and produces actionable insights for the most-reviewed products.She further incorporated Python scripting for precise sorting and formatting.

In her eyes, Python provided control over structure, which matters when downstream decisions rely on it. As agents scale into analytical workflows, deterministic layers become more important, not less. Structured data sorting, explicit package control, and transparent formatting give teams confidence in outputs.

She also highlighted something experienced builders know well: human-in-the-loop is not a fallback. It’s an accelerator. Review windows, clear reasoning summaries, and staged approvals build trust quickly. Once confidence is earned, automation can expand. See her AI agent live on Automation Anywhere Agentic App Store.

🥈 Second Place: Omkar Mahajan – A Transit System That Thinks Ahead

Omkar Mahajan designed his AI agent around a familiar operational scenario: incident response and maintenance in a transit system.

When a staff member reports an issue, the agent consults both maintenance manuals and disaster management guides to determine what actions should follow. It can create Salesforce tasks, generate work orders, and even initiate purchase orders when replacement parts are required. Along the way, every action is logged, giving teams clear traceability into what happened and why.

What makes the design compelling is how Omkar explains the role of an AI agent in this process.

He compares traditional automation to a train running on fixed rails. The route is predetermined. If something blocks the track, the system stops. Even if the destination is nearby, there’s no alternative path because the process can only move forward along the scripted route.

An AI agent behaves more like a car navigating with GPS. The destination remains the same, but the path can adapt. If the system encounters a constraint, it can evaluate the situation, identify another route, and continue progressing toward the goal. That flexibility is exactly what operational environments often require.

At the same time, Omkar’s design shows that flexibility works best when paired with clear guardrails. Before creating purchase orders, the agent checks historical ordering patterns. It validates requests against an approved product list. And when decisions fall outside those boundaries, the process can return to established manual workflows.

As agents begin interacting with enterprise systems like Salesforce, these kinds of controls become increasingly important. They allow systems to remain adaptable while still operating within policy and operational expectations.

In a future iteration, he talks about distributing responsibilities across multiple specialised agents. Separating inventory management, purchase workflows, and operational response. He also described connecting the system to a command center so sensors could trigger preventative actions before incidents escalate. Download Omkar’s submission here.

🥇 First Place: Ganesh Bhat – From Sprint Chaos to Structured Execution

Ganesh focused on a workflow many developers and product teams know well: sprint planning.

Whiteboards, emails, transcripts, scattered screenshots, all of it eventually needs to become structured epics, stories, and tasks in Jira. All of these inputs are important, but translating them takes more time than teams often have. His agent ingests unstructured inputs, synthesises planning conversations, checks resource availability, and proposes task assignments, all before asking for human confirmation.

This is where AI agents differentiate themselves from linear automation. They interpret narrative context, reconcile multiple sources, and make dynamic adjustments in response to current constraints.

Ganesh intentionally layered in guardrails in his design. Sensitive data remains within controlled environments. Confidence thresholds trigger human review. Conflicts surface rather than being silently resolved.

As planning becomes more dynamic, adaptive systems reduce coordination overhead. In this case, Ganesh built a goal-based agents that enables team members to spend time on higher value thinking. We’re pretty sure dev teams will love this one available in our Agentic App Store.

Guardrails as Architecture, Not Afterthought

Across all three solutions, a consistent theme emerged: reliability is designed in early:

• Human approval checkpoints
• Approved product lists
• Historical validation
• Data masking for PII and PHI
• Confidence-based escalation

These structural components enable agentic systems to be viable in enterprise environments.

As agents gain autonomy, explicit reasoning paths and staged approvals create operational confidence. Over time, as patterns stabilize, teams can decide where to relax intervention.

This is an area where intentional design pays off. Systems that explain their reasoning are easier to extend, integrate, and trust. And that’s what makes experimentation sustainable.

Community as a Design Multiplier

We mentioned it at the beginning of this newsletter: Pathfinder Community members are impressive storyteller.

Each winner clearly articulated the problem they observed, why an agent was appropriate, and how it was structured. Now that their agents are featured in Automation Anywhere’s Agentic App Store, their clarity makes it much easier for our users to understand how they can use these pre-built agents.

As a reminder, all of the submissions, including these winning agents, are available to download for free to use and test.

For developers in our community, this is our challenge to you: download one of these agents and learn by deconstructing it. Figure out how its built so you can apply the logic to your future projects.

Building agents is serious work. But doing it together as a community makes it fun.

Dive Deeper

If you want to see how these winners approached their ideas and hear directly from them about their design decisions, we hosted them in this episode of the Agentic Edge podcast.

Watch the full episode on YouTube or listen on your favorite audio platform.

Never miss an episode: Watch on YouTube | Subscribe on Spotify | Subscribe on Apple Podcasts

And if you’re an automation developer looking to stretch beyond task automation into goal-based systems, consider following Automation Anywhere Pathfinder on LinkedIn to stay close to upcoming challenges, meetups, and hands-on opportunities to build alongside peers in automation space.