I gave a talk on autonomous agents at the European Power Platform Conference in Vienna recently and thought I’d share key reflections and the slide deck here. In my talk I also showed a reasoning agent we have at Advania UK which uses our client and Microsoft licensing/SKU data to derive complex technology plans and draft proposals (built with Copilot Studio) - we'd love to show this to you if you're interested in agents.
In general, there’s a mix of hype/reality in agent autonomy today but it’s 100% where we’re going. Before I go into the recommendations, let's define autonomy a little bit - here's a slide from my talk:
In short, we're not defining what the agent does step-by-step - the agent is more outcome-driven. An autonomous AI agent is able to reason and use tools and data we supply it with. There are lots of guardrails of course, and ultimately a few key techniques to understand and apply when building agents of this type.
The link to the deck is at the end of this post, but here's a summary of key points.
Takeaways
1. Invalid expectations 🚫
From the demos, it often seems as easy as creating an agent with instructions like: “Respond to each lead in CRM, send a personalised e-mail and update the event tracker” – job done! However, there’s far too much ambiguity there for any system, even with gen AI. All solvable, but I see people approaching agents and expecting magic.
2. Agents need help 🤖
Related, expect to build out things your agents can use – Agent Flows in the Microsoft world, usable data, custom connectors and so on. If you’re automating documents (e.g. generating proposals), expect to lean on wider platform capabilities like M365, SharePoint and Syntex. In many ways, this is the work with agents – and you need expertise in wider AI, data, automations, systems integration, and DevOps to be effective here. The “death of the developer” is overblown, even if the focus changes.
3. Some key techniques are critical 📚
I listed 5 critical techniques in my talk – ranging from how to ensure an autonomous agent chooses data and tools correctly in its processing, to writing effective agent instructions. Example: often you actually don’t want an autonomous agent to “reason the process” - things go off the rails and the agent gets confused. Defining a process explicitly with reasoning in some steps is better.
4. Understanding governance and cost considerations is vital 💹
Autonomous AI costs money - in Microsoft, ServiceNow, AWS, Google or anywhere else. If you can’t forecast costs you could be making some bad choices. Further, laying down agent guardrails in data access and connectors, agent sharing, identity use, controlled zones for organisation/policy etc. is vital.
5. Conclusion - balancing hype vs. reality ⚖️
In summary, lots of hype and certain pitfalls, but also potential for huge value creation in today’s era of reasoning models and new protocols. I’ll expand on the above in articles on my blog soon, but suffice to say this is the next generation of solutions we’ll build in the next few years.
Link to the slide deck
