Insights on Agentic AI, Machine Learning, and Business Transformation

Washington demanded Anthropic switch off its best model for non-US users. But the AI stack offers surprising ways to maintain sovereignty away from AI models

Anthropic measured something like emotions inside Claude — not mimicry, but representations that direct it. Intervene and the behaviour changes. The debate fixates on one question: is AI conscious? This suggests both sides ask the wrong thing. What does it mean for business?

Gartner says a third of enterprise software purchases will involve an AI agent by 2028, and machines are assumed immune to persuasion. We ran 8,000 trials across five frontier models. Which techniques work, which backfire, and what does that mean for selling to agents?

The gap between top models has collapsed to 5%, and GPT-3.5-level inference cost fell 280-fold in under two years. If AI is now a commodity, competitors can buy the same agent tomorrow. So where does durable advantage live, and why can't it be bought?

When cognitive labour was scarce, firms built processes to protect it. Now agents generate drafts and fixes faster than humans can inspect, and review becomes the bottleneck. If intelligence is suddenly cheap, what becomes scarce inside the firm, and what must leaders redesign?

METR's data shows the length of tasks AI completes reliably doubling every seven months; humans hold focus for twenty minutes. Attention, scarce for fifty years, is becoming abundant. If attention is no longer the constraint, what is, and where should you deploy agents?

Amazon, Walmart, Shopify and Alibaba are deploying shopping agents, and 2025 quietly settled the standards letting them transact at scale. Brands optimised for clicks now must optimise for agents. So who owns the customer when an agent stands between you and them?

Once agents start finding, negotiating and buying on our behalf, they become participants and the economy shifts. Drawing on MIT and Harvard research, this maps ten templates where new players and markets are emerging. The harder question: which is coming for your business first?

A new MIT and Harvard report reaches back to Coase's 1937 theory of the firm and Gale-Shapley matching to argue the real prize isn't automating old processes. The boundaries between firms and markets start to move. So where does that leave your competitive advantage?

Berkeley's second Agentic AI Summit drew Google, OpenAI, NVIDIA, IBM and frontier researchers for talks on where agents are heading, from Chi Wang's MassGen to the Linux Foundation's 'Internet of Agents'. So what does the near future of agentic AI actually hold for enterprises?

Recruiters drown candidates in AI filters; candidates fire back with AI-written CVs. Every plug-in escalates the arms race. But two agent standards, MCP and A2A, quietly change the dynamics. So what would hiring look like if redesigned for a world where everyone has AI?

In 2025 Google released the Agent Development Kit and Agent-to-Agent protocol, the plumbing for assistants that search, compare and transact for users. Soon your customer isn't a person clicking a link; it's an agent. So how do you keep your service in the running?

Google's AI Co-Scientist matched in two days what PhD researchers took months to discover on bacterial gene transfer, and published how its agents did it. The architecture shows where real AI value comes from. So what separates the firms that profit from those left behind?

With Simon Torrance we mapped the agentic AI ecosystem into a three-tier stack: engagement, capabilities and data. Some layers teem with innovators; others are near-empty. But the foundational layer isn't a technology at all. So which capabilities should an enterprise in-source, and which buy in?

Buying agentic AI from one hyperscaler looks convenient. But agents encode your business's intelligence, so handing the lot to Microsoft or Google risks core capability. Firms in banking, insurance and pharma see the danger. How do you gain without becoming a 'Business as a Service'?

Agentico is committing one day of staff time a week to Apart Lab's AI safety and interpretability research. With models like OpenAI o1 now matching physicians on reasoning, advising on AI has become less like automation engineering and more like recruiting.

At a Goodfire and Apart hackathon, our team built a tool to catch hallucinations in medical diagnostic AI, then steer the model around them neuron by neuron via a simple API. What does it mean when a business can edit a model's internals this easily?

Agentic AI flips users from doers into managers of agent teams, yet the 'AI native' interface is still an unwritten page in the UX handbook. Tools from OpenAI, Cassidy and Zapier get the workflows right, but leave users quietly disappointed. What are designers missing?

OpenAI has floated agents priced like a wage at $2,000/month, but if agents aren't people, why pay a salary rather than for results? Moving from SaaS subscriptions to variable, task-based pricing turns every employee into a manager of contractors. Are businesses ready to govern it?

Strip away the hype and AI agents are just LLMs paired with tools and a layer enforcing your rules. From document processing to lead funnels and RFP responses, we set out five paths to value with workflows to evaluate today. Where should you start?

Gartner has named agentic AI its top strategic technology trend for 2025, predicting 15% of day-to-day work decisions will be made autonomously by 2028, up from 0% today. The open question for leaders: which gains are genuinely feasible now, not in 2028?

Everyone argues about AI writing content. The bigger shift is AI now reading it: businesses deploy agent teams to search, rank and summarise thousands of pages, ignoring clickbait and pretty design. So what must marketers actually change when their audience is a machine?

You have thousands of contacts across LinkedIn and your CRM. Each week a handful quietly signal they are ready to do business, but you cannot read them all. Agentico's AI agents sift the noise for you. So how many are ready to buy right now?

Google DeepMind quietly released GemmaScope, tools that expose and tune individual 'features' inside a language model. Agentico used them on Gemma-2-9B to change how an AI reasons, even who it thinks it is. Why does its 'civil law' feature switch off when you type 'Putin'?

OpenAI's o1 and Anthropic's Claude 3.5 both score just 21% on the ARC reasoning benchmark, against 84% for humans, with a $1m prize. We wrapped Claude in an agentic framework that forms hypotheses, tests them and course-corrects. How far did the score jump?

Today's students face 40-year careers managing AI agents that write code, support customers, even hire people. Researcher Ajeya Cotra's thought experiment, an orphaned eight-year-old hiring an adult to run a $1 trillion company, shows why that is harder than it sounds. So what should they study?

A year ago few had heard of agentic AI. Now Nvidia's Jensen Huang calls the opportunity 'gigantic' and 96% of executives told Accenture it is significant. Venture capital sees agents unseating SaaS. So where does that leave the software your organisation already runs on?

Meta has spent billions on AI agents, yet no one has set out a coherent vision for where they are driving us. So we distilled one from the public statements of Sierra, Microsoft, DeepMind and dozens of others. What is agentic AI actually for?

On 20 June 2024 two coding agents shipped together, Builder.io's Micro Agent and Anthropic's Claude 3.5 Sonnet, while Together.ai's Mixture of Agents quietly beat GPT-4. Why do all three foreground a thinking process and testing, and what does that pattern mean for every business workflow?

The architecture behind ChatGPT was invented to translate — and that is the clue to defensible products. Pair an LLM's natural-language interface with machine learning's specialist analysis. But how do you tame hallucinations and industrialise it for thousands of users?

Human-in-the-loop oversight doesn't scale — people suffer 'vigilance decrement', and nobody could review the thousands of lines of code GPT-5 writes daily. Meanwhile AI safety researchers build techniques where systems check each other. So how do these methods become a commercial edge?

A pleasure to join the Tech and Innovation roundtable on 5 June, led by Oxford Brookes University — a room of R&D-focused SMEs, from carbon fibre manufacturers to zero-emission transport operators, where we chipped in for AI. Thanks to Richard Rosser at B4 Business.

In one week Google launched its Vertex AI agent framework and Andrew Ng devoted The Batch to multi-agent collaboration. So why does splitting a task across 'software engineer', 'QA' and 'product manager' agents outperform a single agent — and how should leaders manage them?

Our debut at the Shropshire Business Festival on 11 April landed the same day as Google's AI Agents keynote, sparking lively discussion on generative AI and agents in the workplace. Thanks to Ben Simpson, Hugh Strickland and Chris Henderson.

Office chatbots like ChatGPT breed an unmanaged 'black market' of shadow IT. Wharton's Ethan Mollick argues AI agents are different. So why can agents act where chatbots only advise — and where do today's models still fall short?

Across dozens of AI strategy sessions with SMEs, the same themes recur. The 'AI is hype' versus 'AI will take over' debate distracts staff from the opportunity. Everyone has the same AI, so capability is no edge. So where does an SME's advantage come from?

Microsoft Research gave our data science team a shout-out at their annual conference for a team of AI agents, built on the AutoGen framework, that converse freely to write code and investigate data in Jupyter Notebooks.

Google Brain founder Andrew Ng says agents are the trend to watch. He sets out four reasoning patterns — reflection, tool use, planning and multi-agent collaboration — and a claim: GPT-3.5 arranged into agents can out-code GPT-4. Why would the weaker model win?

Most businesses build 'faster horses' with AI, automating what already exists. Stanford's Erik Brynjolfsson calls this the Turing Trap: machines so good at emulating us they replace us, transferring wealth not creating it. So what are AI's superpowers — and which can't you imagine?

Break agentic AI into its parts — LLM, data, tools, reasoning and environment — and a team of agents can cluster 10,000 apps in a Jupyter notebook with no human prompting. But one ingredient is routinely overlooked. Which part actually makes agents safe to deploy?

Agentic AI arrived in four steps: from DeepMind's AlphaZero, through AutoGPT learning to use tools, to Stanford's Voyager adapting skills in Minecraft, to teams of agents writing and testing their own software. So what made today's agents capable of running a business process?

You become the manager: you set the objective. Three frameworks — AutoGen, MetaGPT and ChatDev — make this real, with agents planning, writing code in Docker and fixing their errors. But LLMs hallucinate. So which tasks can you trust them with?

Microsoft's Autogen lets you hand a team of AI agents an objective and watch them prototype and deliver, echoing Adam Smith's insight that specialised collaboration drove the industrial revolution. But teams bring meetings, groupthink and cost. So when does a team of AIs beat one?