Explainers · Agents

A Visual Primer

It doesn’t just answer.
It acts.

The difference between a chatbot and an agent, and why that difference changes the risk. Seven stages.

Scroll

Answer once, or work until done.

A chatbot takes your message and produces one answer. An agent takes a goal and works towards it: it can take actions, look at the results, and decide what to do next — repeatedly — until the job is done or it gets stuck.

CHATBOT

question→answer

One shot.

AGENT

goal→work→check→work→…→done

Keeps going until the goal is met — or it gets stuck.

Same underlying model. The difference is the loop around it.

Think. Act. Observe. Repeat.

Every agent runs the same cycle. Think: decide the next step. Act: use a tool — search, open a document, run a calculation. Observe: read what came back. Then round again, until the goal is met.

Round and round until the goal is met. That cycle is the whole trick — everything else is which tools it may use and when it must stop.

Watch one run.

A small but real example. Press Step to advance the agent one move at a time and watch the loop from stage 02 play out.

Find the filing deadline in the directions order and add it to chambers’ diary.

THINKI need the directions order. I’ll search the bundle for “directions”.

step 1 of 6

Tools are the hands.

An agent can only act through the tools it is given: search, read files, draft documents, send emails, book entries, run code. Each tool extends what it can do for you — and what it can do wrong. An agent with read-only tools can waste your time; an agent that can send emails can waste your reputation.

searchreaddraftdiaryemailpay

lower riskhigher risk

Capability and risk rise together.

Small errors compound.

A model that is right 95% of the time sounds excellent — but an agent chains many steps. Across 10 dependent steps, 95% per step is roughly a 60% chance of a flawless run; across 30 steps, about 20%. And an early mistake doesn’t stay contained: step 3’s wrong date becomes step 9’s wrong diary entry.

1 step

95%

5 steps

~77%

10 steps

~60%

30 steps

~21%

Chance of an error-free run, at 95% per step. Figures are approximate.

Delegate. Don’t abdicate.

The answer to compounding risk is not to avoid agents — it is to supervise them the way you would supervise anyone acting in your name. Four controls do most of the work.

Checkpoints.

The agent pauses for approval before anything irreversible — sending, filing, paying. It can draft the email; a person presses send.

Bounded tools.

Give the minimum set the task needs. Nothing that spends money or communicates externally unless it is essential to the job.

Audit trail.

Every step logged — each thought, tool call, and result — so you can reconstruct what happened and why, after the fact.

Human review of the output.

An agent’s work product is a draft, exactly like a chatbot’s answer. It gets read before it gets relied on.

The professional-duty position is unchanged: however many steps the agent ran, the work leaves under your name.

From tools to colleagues.

Agents are moving from single tasks to sustained work: reading whole bundles rather than single documents, monitoring deadlines across matters rather than one directions order, operating a computer the way a person does — screen, cursor and all.

The loop stays the same. Think, act, observe, repeat — nothing in stage 02 changes as the work gets longer. What changes is the leash: more steps between checkpoints, more tools, more of the working day delegated in one brief.

So the question for a practice is not whether to use agents but which decisions stay human — and that is a judgment call, made in advance, not discovered afterwards.

The loop is simple. The judgment is yours.

Agents multiply what one professional can supervise — provided the supervision is real. Next: not every job needs the biggest model. How to choose.

Next: Choosing the Right Model →

It doesn’t just answer.It acts.