What agentic AI means when you strip out the marketing

Agentic is doing a lot of work in product announcements right now, and most readers nod at it without being told what it means. It has a real definition, and it is simple: the model takes a job that needs many steps and does the steps itself, using tools, instead of asking you what to do next each time.

The difference is between a model that writes you a function and a model that opens the project, finds the bug, changes four files, runs the tests, and comes back when the tests pass.

What it looks like when it works

The examples in Anthropic's Opus 5.5 announcement are a fair picture of the state of it:

• A 680,000 line code migration finished in under a day by one tester, work Anthropic says would have taken a team weeks.
• Asked to cut load times across every page of a web app, the model succeeded 39 times out of 40.
• A playable game built from a single prompt, scored on graphics and polish rather than on whether it compiled.

Read the second one again, because the interesting number is not 39. It is that the previous model made smaller improvements and changed how the app behaved while making them. Doing the job is one thing. Not breaking something else on the way is the harder part, and it is what separates a demo from a tool.

The three things an agent needs

1. Tools. A way to read files, run commands, search, and see what happened.
2. Memory of the job. Not memory of you. Memory of what it already tried in this task, so it does not go in circles.
3. A stopping condition. Something that tells it the job is done or that it is stuck, rather than working forever.

Most agent failures you will meet are the third one. The model does not know it has failed, so it keeps going confidently in the wrong direction, and the bill keeps running with it.

The risk that comes with the capability

If a model can act, then anything it reads can try to give it instructions. That is prompt injection: a web page or a document containing text aimed at the model rather than at you.

Anthropic says Opus 5.5 is more resistant to prompt injection than Opus 5 and is less likely to take actions that are hard to undo. Both of those are the right things to measure. Neither is a solved problem, and any vendor claiming it is solved is telling you something about themselves.

Should you use one yet

For contained jobs with a clear finish line, yes, and the time saved is real. For anything touching money, production systems or data you cannot restore, keep a person between the model and the button. Not because the models are bad, but because the cost of the rare failure is not shaped like the cost of the common success.

Tools that run in your browser, no agent required

Where agents still fail

Worth knowing before you hand one a real job, and none of these are exotic.

1. Confident wrong direction. It decides the problem is in one place, and every subsequent step builds on that. Tests catch this. Watching it does not, because it sounds reasonable throughout.
2. Silent scope creep. Asked to fix one thing, it tidies four others. Sometimes helpful, sometimes a review nightmare.
3. Loop without progress. Try, fail, try the same thing. Cost runs while nothing moves.
4. Stopping too early, declaring success on a job that is half done, which is the failure you only find later.

The pattern in all four is the same: an agent is bad at knowing what it does not know. Give it a way to check its own work, tests being the obvious one, and most of this improves sharply.

How to give a job to an agent

1. Say what done looks like, in a way that can be checked. Tests pass. The page loads in under two seconds. The file exists and opens.
2. Give the smallest access that allows the job.
3. Work on a copy, or a branch, or anything you can throw away.
4. Read the diff, not the summary. The summary is written by the thing you are checking.

Where this is going, and what stays true

The direction is longer jobs with less supervision, and the releases keep pointing that way. What has not changed, across every generation so far, is that the value comes from bounded jobs with a clear finish line, and the trouble comes from open ended ones near something important.

If you only remember one line: let it do the work, keep the decision.