RPA is dead. Long live AI agents

Plus: a tutorial on building agents with Google's Gemma, a marketplace to sell agents to companies, and more

Author

Michael Cunningham
February 17, 2025

Welcome back to Building AI Agents, your biweekly guide to everything new in the AI agent field!

Also like toddlers, they have incredible contributions to make to the world, but they may make a few messes along the way

In today’s issue…

  • Agents are the death of RPA—and its rebirth

  • A marketplace for agent builders to sell to businesses

  • Turn any LLM into a visual agent

  • How to build agents with Google’s Gemma

  • A leaderboard for the best agent LLMs

…and more

🔍 SPOTLIGHT

Created by the author using Dall-E 3

“Robotic process automation” sounds like the quintessential technology of the future, and it was—until AI agents came along.

In the late 2010s, the term robotic process automation (RPA) rose to prominence in the corporate world, referring to a broad suite of tools intended to automate repetitive white-collar tasks. A typical example of an RPA use-case is invoice processing: a company’s payments department might use an automated workflow to extract vendor, price, and other values from incoming invoices, flag them if they are deemed suspicious according to preset rules, and enter them into an accounting system. Companies such as UiPath, Zapier, Automation Anywhere, and ServiceNow promised to save companies billions by automating significant portions of their workflows with RPA.

By the beginning of the 2020s, however, the field’s promise had begun to fade. Outside of very simple tasks that require the exact same operations every time, most business processes proved to be too complex and require too much subjective human judgement to be done by the kind of rules-based algorithms that RPA provides. Additionally, RPA solutions were extremely brittle: an invoice-processing workflow, for example, could break down completely in the face of invoices with a non-standard format. RPA was increasingly seen as a failed buzzword.

Just as the field seemed to be in terminal decline, however, a new tool came to the rescue: large language models (LLMs). While RPA had made some limited use of natural language processing before the launch of ChatGPT in 2022, that seismic event showed the industry that it finally had the key tool it needed to inject real intelligence into what were previously inflexible workflows. In particular, agentic AI—the use application of LLMs as orchestrators of applications—was seized upon by a host of major tech companies as a revolution in enterprise automation. RPA firms wasted no time in jumping on the agent train.

Automation Anywhere was first, with its CEO Mihir Shukla proclaiming “Automation is the foundation that’s gotten us partway there. But AI Agent-powered automation is the breakthrough that will take us beyond” as the company released a low-code AI Agent Studio to allow users to easily create LLM-powered automations.

Its competitor UiPath declared that the future of its platform is agentic, launching an Autopilot agent that is intended to work across all of a user’s applications, from their computer operating system to Office 365 and more. Following in the footsteps of Automation Anywhere (and many others), it released its own low-code Agent Builder studio, and, underscoring the company’s all-in bet on agents, invested $35 million in French agent startup Holistic.

Automation firm ServiceNow made a similar push into the space, releasing thousands of pre-built agents for tasks such as customer and IT service management and, of course, announcing plans to roll out their own agent studio for users in March.

Zapier, which played a key role in the first wave of RPA by providing integrations with web applications, allowing them to be incorporated into RPA workflows, leveraged its technology to create Zapier Agents, which integrate with over 7,000 applications.

Although some argue that the rise of agents represents the end of RPA, it is better understood as a death and rebirth. While the original tools of RPA proved inadequate to their task, the new generation of agentic AI is finally proving capable of delivering the broad and impactful enterprise automation that RPA initially promised, and RPA firms—all of which now prominently feature AI agents on their websites—live on. RPA may be dead, but AI agents are rising like a phoenix from its ashes.

If you find Building AI Agents valuable, forward this email to a friend or colleague!

🤝 WITH THE AI REPORT

There’s a reason 400,000 professionals read this daily.

Join The AI Report, trusted by 400,000+ professionals at Google, Microsoft, and OpenAI. Get daily insights, tools, and strategies to master practical AI skills that drive results.

📰 NEWS

Source: Perplexity

Rapidly growing AI search startup Perplexity introduced its competitor to OpenAI and Google’s dueling Deep Research offerings.

Thinkers360, a business-to-business (B2B) marketplace provider, launched a directory to connect enterprise buyers with AI agents and agent platforms, allowing agent builders to market their products to companies.

Biotech company Coeptis has launched COEP Venture Group, an investment fund specifically devoted to AI automation and agent startups.

🛠️ USEFUL STUFF

Created by the author using Dall-E 3

The tech giant open-sourced the second generation of its OmniParser, which enables any LLM to navigate graphical user interfaces (GUIs) by converting screenshots into interpretable elements that LLMs can interact with. As a demonstration of its capabilities, Microsoft also released OmniTool, which provides a virtual Windows OS for agents to use.

Google provides a tutorial for building a simple agent with its open-source Gemma 2 model.

Galileo.ai launched a leaderboard for comparing the performance of various commercial and open-source LLMs on agentic tasks, allowing agent builders to identify the most cost-effective models.

💡 ANALYSIS

Source: www.epictop10.com

With Microsoft’s CEO predicting the death of SaaS at the hands of AI agents, this piece makes the case that it will continue, albeit in a more agentic form.

The best way to bring AI agents to an enterprise, the author of this article argues, is to establish a center of excellence (CoE) that can identify and support ideal use cases for them.

This piece weighs into the debate, intensified by Workday’s treatment of agents as digital workers in the launch of its Agent System of Record, over whether agents should be managed by IT like technology or by HR like real workers.

🧪 RESEARCH

A possible attack vector against scientific agents | Source: arXiv

The slate of added tools that give agents their power also makes them uniquely vulnerable to simple attacks, this paper argues, providing a taxonomy of these threats and some illustrative examples against popular agents.

Thanks for reading! Until next time, keep learning and building!

What did you think of today's issue?

Login or Subscribe to participate in polls.

If you have any specific feedback, just reply to this email—we’d love to hear from you

Follow us on X (Twitter), LinkedIn, and Instagram