AI agents unleashed

As the tech giants jostle in the AI investment space — and AI stocks skyrocket — there’s one AI development in particular that’s fast-expanding what AI can do for businesses: the AI agent. 

I like to think of it as ‘any intelligent AI process that can act on behalf of a human to achieve a goal’. The many possibilities that such agents are now opening up have earned them a ‘beginning to see’ place in our Looking Glass ‘AI Trends to Watch’ matrix, where we’re strategically recommending their adoption.

AI agents can be autonomous to varying degrees. For example, you can give an agent a specific goal and then see how well it’s achieved it — like when you ask ChatGPT to find certain information for you, ‘web browser’ style, or use its Python capabilities to interpret data. 

Or you can ask an AI agent to come up with a multi-step plan to achieve a more complex goal — which it can systematically work through with your approval. Or you can go further, beyond traditional robotic process automation (RPA) thinking, to support ongoing business processes by using corporate AI agents that are now far more adaptable and intelligent.

Thoughtworks’ project modernizing fragmented systems for a multi-play Canadian telecoms company highlights the potential of this last route. Rather than a human manually deprovisioning a service and re-commissioning it to address a customer’s problem, a fully autonomous AI agent could instead identify the issue and then solve it in the background. A semi-autonomous approach might involve a customer service rep identifying a problem and then instructing the AI agent to implement a suggested solution.

But it’s perhaps an AI agent’s ability to interface with corporate systems and real-world data via APIs that, for me, is one of the most interesting current developments. OpenAI’s GPT models, paired with Zapier to connect to 6,000+ corporate systems including Trello and Jira, is one example, and there are others such as Amazon’s Bedrock platform and Google’s Duet AI.

For example, while AWS Bedrock allows you to review an agent action plan and its ‘thought’ processes, enabling you to check they are what you’d expect them to be, reliably testing AI agents in general remains an industry challenge. In the end, how much time and effort you commit to this will be influenced by the level of risk involved. While leaving a personalized marketing email process to run automatically may not have serious implications — especially when 95%+ of those emails sent will remain unread — the testing burden becomes much higher should the AI agent be interfacing with, say, an accounts receivable API. 

To help with this we’re seeing significant investment increases from the big AI players to improve reliability, safety and testing. All this is underpinned by greater accountability and transparency as the capabilities of AI agents look set to increase. Alongside an ever-sharper focus on reliability, risk and capability, companies will also be looking to bear down on costs as they scale the use of agents, which can demand significant compute resources behind the scenes.

AI agents mark a step change in how companies can harness the power of AI interacting in richer ways with its environment to achieve their business goals faster, more cheaply and more effectively. And while fully recognizing the concerns of some around AI agents’ autonomy, at Thoughtworks we believe that maintaining a laser-like focus on their management, direction and output is the best way for companies to reap the benefits of their huge potential.

If you're interested in AI agent or other emerging trends in the field of AI, why not check our Thoughtworks’ latest Looking Glass report? It's packed with useful insights to enable you navigate the next stage of your AI journey, along with other important tech trends.