Agentic AI is everywhere – the portents of its revelatory prowess are on the lips of not only tech evangelists, but even stoney-faced cynics who are yet to buy in to the idea that the ROI on AI is anywhere near commensurate with the resources being globally rained upon it.

Whatever your take – there’s some very compelling data around AI agents as a tangibly ROI-savvy application of the technology.

As of last year, less than 1% of enterprise software applications included some form of agentic AI, but tech consultancy Gartner has predicted that by 2028 that number will rise to 33%, allowing for 15% of day-to-day work decisions to be made autonomously.

While previous AI models were limited to generating text or summarising interactions, agentic AI introduces a new paradigm where AI systems possess the capability to act autonomously to complete tasks.

So, what does the world’s biggest AI firm have to say about it?

In his keynote speech at Turing Fest 2025, Jonathan Field, solution architect at OpenAI, discussed OpenAI’s recent Agent releases, illustrating how these new capabilities can enhance real-world business applications.

Field also discussed the key factors behind successful outcomes when implementing agentic AI, common reasons why initiatives stall, and clear, practical guidance on how businesses can navigate this transition without getting caught in typical – and wildly expensive – pitfalls.

What is Agentic AI, why is it Important, and What is OpenAI Doing in this space?

Field opened his keynote with a stark observation: “You can’t go on LinkedIn without seeing somebody mention [agentic AI]… The reason everybody is talking about it is because there’s a huge shift going on right now.”

He contrasted today’s AI with early ChatGPT: “When ChatGPT first came out, everybody got used to this term-by-term conversation. You ask a question, you get the answer. The shift now is moving away from AI doing work with you to AI doing work for you – and that’s quite a big shift.”

OpenAI’s recent releases are betting big on the power of agentic AI – tools designed to handle complex, multi-step tasks with minimal supervision. And all with that user friendliness and low barrier for entry that’s made the company proliferate so intensely since November, 2022.

One such tool, Deep Research, showcases how AI can autonomously conduct time-intensive processes like competitor analysis, market landscaping, or policy benchmarking (among a million other potential use-cases). Rather than simply surfacing search results as the standard chatbot will do, it works across a “longer horizon” – identifying relevant sources, reading and comparing documents, extracting key findings, and delivering structured insights.

In contrast to traditional assistants that require constant prompting, Deep Research operates with a degree of autonomy that begins to resemble a competence and thoroughness that you could (anecdotally) equate to a junior analysts work, capable of adapting as it uncovers new information or shifts in context – so, you know… agency.

Another recent release from OpenAI, Operator, works as a “computer-using agent” designed to automate repetitive web-based tasks by interacting with software in much the same way a human would – according to Field.

Built into ChatGPT, Operator can navigate websites, fill out forms, click buttons, and simulate complex workflows – all while responding dynamically to visual and contextual cues on a page.

Field highlighted its enterprise potential, saying: “We’ve seen customers using it to automate usability testing on websites… rather than a human having to go through and check everything.”

This opens up possibilities for streamlining quality assurance, customer journey testing, and other routine digital operations that typically demand manual effort.

Practical Advice for Businesses Implementing Agentic AI

Field didn’t just outline the potential of agentic AI and use the platform as a product pitch for OpenAI – let’s be honest, the company doesn’t really need the leg-up.

In his talk, Field got into the weeds of the how, offering blunt, refreshingly tangible advice for businesses navigating this shift to agentic AI.

His guidance revolved broadly around three pillars: scoping, simplicity, and relentless measurement – with some off-shots here and there that we’ll go through below:

1. Scoping: “Do You Even Need an Agent?”

Before writing a single line of code, Field urged businesses to ask a foundational question: “You may not actually need to build a logic system… You might be able to do it in a simpler way.”

He compared the current hype to 2024’s rush to custom models: “Everybody was absolutely convinced they needed to build their own model… Actually, a lot of the functionality is already there.”

His litmus test? “If you have a use case where something works 80% of the way and that’s okay, you should just do that… Every time you write code, you commit to maintaining it.”

2. Design for Simplicity: Single-Agent vs. Multi-Agent Systems

Field repeatedly emphasised starting with the least complex solution. Most businesses, he argued, overestimate the need for multi-agent setups: “Most people see a tonne of value from single-agent systems… It’s very good if you have quite narrow domains.”

He illustrated this with an example: “If you’re building a Slack bot to pull sales forecasts or update a Google Sheet, a single agent works really well.” Multi-agent systems, while powerful, introduce unnecessary risk for many use cases: “If you have multiple agents handing off to each other, there are so many things that could go wrong—tools failing, instructions getting confused.”

But how do you know when to scale up? Field’s rule of thumb: “If you’ve got diverse instructions, lots of tools, or high-stakes consequences for mistakes, then consider breaking the problem into specialised agents.”

3. “Do the Hardening First”: Pre-Launch Checks

One of Field’s most emphatic warnings centred on pre-launch groundwork – a step many skip: “Does the underlying system you’re using have the APIs you need? Can it handle production load? If you’re hammering it with 2,000 requests a second, will it fall over?”

He stressed aligning with subject-matter experts early: “For a customer service agent, you need humans to say, ‘This is how we do this.’ Unless you understand the business process, you won’t be successful.”

4. Iterate Hard, Measure Harder: The “Evals” Mandate

Field called evaluations (evals) the “single most critical thing” businesses neglect. His definition was succinct: “It’s like unit testing for generative AI… a systematic way to tell if the agent is doing the right thing.”

He outlined a pragmatic approach:

• Start small: “Begin with 30–50 test cases. For customer service, grade each output: Did it mention banned terms? Did it follow instructions?”
• Use frameworks: “Don’t build your own eval system unless you’ve got a compelling reason. There are plenty of open-source options.”
• Beware of “vibe-based” testing: “I’ve seen teams spend five months coding, then evaluate performance based on vibes. Not good.”

Field also shared a troubleshooting framework: “If the agent isn’t doing what you wanted, it’s usually one of two problems: a context issue (it lacked information) or an instruction-following issue (it didn’t do what you asked). Classify it, then optimise.

Recommended reading

• 40% of Tech Pros Expect to Make at Least 3 Career Changes
• Decline in UK Tech Job Adverts Threatens Young Talent Pipeline
• What Will Be Scotland’s Best Paid Tech Jobs in 2025

Field displayed this live by showing a deep-fried Mars Bar – with the AI classifying the item as “Scottish cuisine.” When pressed as to how it came to this conclusion, the AI displays that through its research, it found that this gastronomic delight came to prominence through Scottish chip shops, hence the leap to “Scottish cuisine”.

5. Observability: “Pinpoint Where Things Broke”

For multi-agent systems, Field stressed the need for observability tools: “If a tool fails or a handoff goes wrong, you need to know where.”

He highlighted OpenAI’s SDK and open-source options: “They show you each stage of the agent’s journey. If something went wrong, you can pinpoint it fast.”

6. Start Low-Risk, Then Scale

Field’s final piece of advice relates to the early days of any major tool implementation into your tech stack: “Your risk tolerance early on is low. Start with high-volume, low-value tasks where mistakes aren’t catastrophic. Prove it works, then move up.”

He warned against overreaching: “Don’t go straight for the most complicated task. I’ve seen teams build three different agents for the same thing when one would’ve sufficed.”

While Field’s keynote covered vast ground, two principles stood out:

1. Design for Simplicity: “Use the simplest solution that works. Single-agent systems solve most problems.”
2. Iterate Hard, Measure Hard: “If you skip evals, you’re flying blind. Every model update could break your agent – and you won’t even know.”

Calling agentic AI a buzzword – as many do – is quickly becoming a very hard position to defend. However, that doesn’t mean there isn’t a whole lot a fluff and false dawns around the technology.

Field’s keynote commendably cut through the hype: start small, keep it simple, and measure everything.

As AI shifts from assistant to operator, those who focus on practical and measured execution – not just flashy demos and charitably framed metrics – will be the ones who actually see results.