Watch an Agent Use Tools

It can be surprisingly satisfying to watch an AI agent work: searching a document, checking a result, creating a draft and then coming back with something usable. But how do you tell genuine progress from a stream of impressive-looking activity?

Imagine an agent preparing a meeting briefing. It might search authorised notes, identify decisions made last time and create a draft agenda. Each tool call does a specific piece of work.

A good progress report tells you what's completed, what remains uncertain and which actions are awaiting your decision. 'I'm thinking carefully' says very little; 'I found the minutes and drafted three agenda items' tells you something useful.

Some work can be automated fully within the agreed boundaries. Other steps, such as sending an external message, may need an explicit approval. This isn't necessarily a weakness: it lets an agent handle preparation while you retain decisions with real-world consequences.

As you become familiar with these stages, agents stop feeling like mysterious autonomous beings. They're systems using particular tools, with observable results and limits you can understand.

Watching an agent work can be fascinating. It searches a website, opens a document, extracts a figure, creates a file and then searches again. But a stream of activity isn't the same as a successful result. What matters is whether those actions moved the job towards the outcome you asked for.

Imagine you've asked a connected assistant to prepare an agenda for a community group's monthly meeting. It searches the shared folder for previous minutes, opens an actions spreadsheet and drafts a new document. Those are tool calls: the AI-directed process has requested specific operations from other software rather than pretending to have done everything through text alone.

The first search might return three files with almost identical names. The assistant has to select the correct meeting period, or ask for clarification. If it chooses the wrong file, everything after that may look polished while relying on outdated actions. Tool results provide data, but they don't automatically resolve ambiguous context.

Now suppose the spreadsheet service is unavailable. A poor system might continue and announce the agenda is finished, leaving out the outstanding actions. A better one reports that it completed the minutes summary but could not retrieve the latest action register. That distinction between attempted, completed and blocked work is essential when AI starts doing things for us.

Sometimes a tool call can have consequences beyond reading information. Creating a document changes the state of an external application. Sending the agenda changes what other people receive. You may be happy for the assistant to create a draft but still want to review it before anyone gets an email. The agent needs to know those boundaries.

Ask for a concise activity record at the end: which source files were used, which draft was created, which steps failed and what still needs your attention. Not every click needs to become an audit report, but you should be able to distinguish a genuine completed artefact from an assistant's description of what it intended to create.

The most useful way to assess an agent is therefore outcome-first. Did it produce the correct agenda, using the latest information, in the right place? Could you inspect and edit it? Did it tell you about failures? A system that performs fewer actions reliably may be more valuable than one that looks impressive while endlessly navigating.

Now imagine the agenda document is created correctly, but in the wrong folder. Technically a tool completed its operation, yet the job isn't finished because the group cannot find the result. A useful end-to-end check asks whether the artefact exists where it should, can be accessed by the intended person and contains the agreed information.

There's a distinction between observing an agent's intermediate steps and controlling every step. You may not need to inspect each search query. You do need meaningful checkpoints: the sources selected, the output produced and any consequential action awaiting approval. Well-designed visibility helps you supervise a process without becoming its full-time operator.

This idea becomes even more important when several tools are connected. A system might successfully read from a drive but fail to create a document. It should not claim the whole workflow succeeded simply because one stage completed. Asking for explicit status at each handover makes failures easier to identify and correct.