Agents, and why 2026 is different

Everything in the earlier parts describes a tool that answers: you ask, it writes, you decide what to do with the text.

Through 2025 and 2026 the products have been growing a second mode, and the vendors' word for it is "agentic".

It is worth understanding, because it changes the risk calculus more than any model upgrade has.

A chatbot answers; an agent acts

An agent is the same prediction engine given two things a chatbot does not have: a goal, and tools it may use on its own initiative — searching, opening documents, running code, filling forms, sending the email, filing the document. Instead of producing one reply, it works in a loop: assess the situation, take a step, look at the result, take the next step, until it judges the goal met.

Nothing about the engine changed. The hallucination mechanism, the confident style, the agreeableness — everything in part two still applies.

What changed is what happens next: a chatbot's mistake sits in a draft until a human reads it; an agent's mistake can be acted on by the agent itself, in the world, before anyone looks.

Why the risk compounds

Three properties make agent mistakes a different category.

Errors chain. An agent's steps feed each other, so one wrong judgment early — the wrong document opened, the wrong name extracted — propagates through everything built on it. A 95%-reliable step run twenty times in sequence becomes a coin flip.

The review point moves. With a chatbot, human review sits naturally between the tool and the world: the model drafts, you read, you send. An agent collapses that gap. Unless a checkpoint is deliberately built in, the natural moment to catch the error is gone.

The consequences are real. A hallucinated paragraph wastes your time. An agent that emails the wrong attachment to the wrong recipient, or agrees something on your behalf, has done something that cannot be undone. The question to ask of any agent is not "how good is the model?" but "what is the worst single action this thing can take without me?"

Where adoption actually is

This is moving quickly, so date everything.

The Thomson Reuters Institute's 2026 professional-services survey (February 2026) put agentic-AI adoption at 15% of organisations, with a further 53% planning or considering it. The assist-to-act shift is arriving into the same strategy-and-measurement vacuum described in the essay this module opens with.

The first regulator has spoken. The Financial Reporting Council — the UK audit regulator — issued guidance on generative and agentic AI in March 2026, describing it as the first from any audit regulator globally. Its position is instructive: it supports firms realising the benefits, and insists that "it is people … who are accountable for audit quality — this guidance does not alter that position." Expect that structure — use permitted, accountability undelegable — to be the pattern as other regulators follow.

The posture that follows

For agents, the discipline from the earlier parts sharpens into one rule: a human checkpoint before anything leaves. Let an agent search, read, assemble and draft — that is the same supervised work as parts one to three, at greater speed. The line worth holding is at actions that touch the world: sending, filing, signing, paying, publishing. Whether and when to delegate those is the kind of deliberate, understood decision this module exists to make possible: a bearing to choose, not a default to drift into.