Frontier AI: reasoning models and agents

0:00 Frontier AI

If you last looked closely at AI a couple of years ago, two things have changed dramatically. Models now reason, spending extra effort thinking before they answer. And they act, using tools and working through many steps towards a goal. In this lesson you'll understand both shifts, where they genuinely help, where they still fail, and what's worth watching next.

0:26 Reasoning models

Reasoning models first. OpenAI's o1, in September 2024, popularised training models to work through an extended chain of reasoning before giving a final answer. In January 2025, DeepSeek's R1 showed strong reasoning could be released with open weights, which intensified competition and pushed prices down. Today most major model families offer a reasoning or extended thinking mode, often with a dial for how much thinking to allow.

0:55 Why it matters

Why does this matter for your business? Because these two shifts change which tasks AI can take on. Two years ago, many teams tried AI on analysis and planning and were disappointed by shallow answers. Reasoning models changed that for many analytical tasks. And agents moved AI from drafting text to doing multi-step work in real systems. That means the list of things worth trying has grown, but so has the need for careful scoping, verification and access control. The organisations getting value are the ones that understand both the capability and its limits.

1:36 Corridor, notepad, laptop

Here's an analogy. A fast model is like asking a knowledgeable colleague a quick question in the corridor. A reasoning model is like asking them to sit down with a notepad for twenty minutes. For where's the meeting room, the corridor answer is perfect. For should we restructure our pricing, you want the notepad. And an agent is that colleague with a laptop, a phone and access to your systems, who can go and do the work, which is exactly why you give them clear instructions and check their output.

2:15 Simple example: two quotes

A simple example. You ask a fast model to compare two supplier quotes. It says supplier A is cheaper. You ask a reasoning model with the same quotes. It works through delivery charges, payment terms and a minimum order quantity buried in the small print, and concludes supplier B is cheaper for your actual volumes. Same question, same documents, different depth. That's test-time compute earning its cost, and it's why you check the reasoning, not just the conclusion.

2:49 Test-time compute

The idea to remember is test-time compute. You can buy a better answer on a hard problem by letting the model think longer, at the cost of time and money. That gives you a new design choice. Fast, cheap responses for routine work like rewriting an email. Slower, deeper reasoning for high-stakes analysis like comparing contracts or explaining a financial variance.

3:16 Strengths and blind spot

Reasoning shines on structured analysis, coding and debugging, planning, maths and logic, and critiquing its own drafts. But it has a blind spot. Reasoning can't fix wrong premises. Give a model an outdated price list and it will reason brilliantly to the wrong conclusion. It still needs current, correct information, through documents you provide or retrieval from trusted sources.

3:42 Agents

Now agents. An agent combines a model with tools, like search, code, APIs and browsers, with memory of what it has done, and with a loop: plan, act, observe, adjust. The Model Context Protocol made connecting models to tools and data much easier, and agent-to-agent protocols like A2A aim to let agents from different vendors work together. In short, AI has moved from answering questions to doing work.

4:12 Where agents deliver

Where are agents delivering real value today? In software engineering, coding agents read codebases, make multi-file changes, run tests and open pull requests for developers to review. In research, deep-research modes search, read and synthesise many sources. In operations, agents triage tickets, reconcile data and draft responses, with humans approving anything consequential. The pattern is consistent: well-scoped tasks, clear success criteria, good tools and human oversight.

4:41 Still risky

And where are they still risky? Open-ended autonomy on high-stakes tasks. Errors compound across many steps, and agents can be manipulated by content they read, a problem called prompt injection. That's why sensible teams scope tightly, restrict access, and keep humans approving actions that matter. Treat any claim of fully autonomous everything with healthy scepticism.

5:05 What to watch

What should you watch, as outlook rather than prediction? First, how long and complex a task agents can complete reliably, tracked by independent evaluations, not launch videos. If that keeps improving, more multi-hour knowledge work becomes delegable. Second, cost per unit of capability, which has historically fallen quickly; re-cost your shelved ideas every six months. Third, governance: the EU AI Act, sector regulators and standards like ISO 42001 will shape how agents can be deployed. Here's a real-world style example: a Dubai trading company uses reasoning mode for month-end variance analysis, but every explanation must cite ledger lines a human checks, and posting journals stays human-only.

5:51 When to use reasoning mode

A practical way to decide when to use reasoning mode is to ask two questions. Is the task hard, meaning multi-step, analytical or easy to get subtly wrong? And are the stakes high enough that a better answer is worth waiting and paying for? If both answers are yes, use reasoning. If either is no, a fast model is usually fine. Write that rule down for your team, and review it when new models arrive.

6:24 Three mistakes

Three common mistakes. First, using reasoning mode for everything, which is slow and expensive, or for nothing, which misses its value on hard problems. Second, trusting a long, confident chain of reasoning without checking its premises. Third, giving agents broad access before you've scoped the task and tested it. The safest path is narrow tasks, clear success criteria, least-privilege access, and humans approving anything consequential.

6:52 Try this now

Try this now. Pick three real tasks from your week: one routine, like rewriting an email; one analytical, like comparing two offers; and one planning task, like a ninety-day rollout. Run each in a fast mode and in a reasoning mode of the assistant you already use. Check the answers yourself for accuracy, note how long each took, and rate usefulness from one to five. Then write one sentence: we use reasoning mode when, and fast mode when. Pin that sentence where your team will see it, and revisit it when new models arrive.

7:33 Recap

To recap. Reasoning models trade time and cost for better answers on hard problems, but still need correct inputs. Agents add tools, memory and a loop, and work best when tightly scoped and overseen. Watch reliability, cost and governance using independent evidence. Your next step: run the reasoning-versus-fast test in the lesson text on three real tasks, and write one rule for when to use each mode. Next, multimodal AI, world models and on-device AI.