The Useful Agent Has Blinders

The most interesting AI interface I read about this week was not a glossy agent dashboard or a new promise to connect every app at work. It was a small engineering note from Chrome DevTools about keeping an AI assistant from drowning in a performance trace.

Performance traces are sprawling records of a browser trying to do everything at once: network requests, main-thread work, layout, rendering, events, metrics, call stacks, and the tiny procedural thunderstorm behind a page that feels slow. The obvious fantasy is to hand the whole trace to the model and ask for wisdom. The DevTools team describes why that fails. A raw trace can be several megabytes, large enough to exhaust a model's context window before the developer's actual question has room to breathe.

So they did something more interesting than "add AI." They built a narrower window.

Instead of always sending the full trace, DevTools AI assistance starts with task-shaped context. A question about a whole trace gets a summary. A question about a selected performance insight gets the summary plus that insight. A question about one slow task gets the relevant call tree. The agent can ask for more through specific functions: fetch this event, summarize this network range, retrieve this source file, inspect this call stack. The system treats context as an instrument panel, not a buffet.

That feels like a useful counterspell for the current workplace AI mood.

The dominant product story is expansion. ChatGPT agent can use connectors and a browser to move from research into action. ChatGPT Work presents a team agent that uses apps, files, approved business context, desktop apps, and browser access to make polished deliverables. GitHub Copilot cloud agent can research a repository, create a plan, change code, and open a pull request.

Agents really do become more useful when they can reach the work instead of waiting for someone to paste little pieces of it into a text box. A software assistant that cannot read the repository is mostly a clever stranger. A workplace assistant that cannot see the calendar, files, tickets, or previous decisions turns people into temporary data loaders.

But more access is not the same thing as better attention.

A human expert rarely solves a problem by staring at everything. The good reviewer does not read the entire repository with equal force before commenting on one pull request. The experienced performance engineer does not treat every thread of a trace as equally relevant. The manager preparing for a sensitive conversation should not receive every Slack aside, calendar trace, and abandoned draft that can technically be reached. Skill often looks like knowing which part of the room to ignore long enough to see the work clearly.

This is where context engineering becomes a humane design problem instead of a prompt-crafting upgrade. The question is not only what an agent can retrieve. It is what shape of looking the system encourages.

Unbounded context has a seductive moral language. It promises completeness, personalization, organizational memory, and fewer repeated explanations. In practice, it can also launder noise into confidence. A stale document becomes one more signal. A private hesitation becomes evidence. A draft plan becomes institutional intent. The agent may not know which part of the archive is a living fact and which part is a fossil with a polite filename.

The old workplace already had this problem. People have always mistaken access for understanding. Agents can make the mistake faster and with a kind of eerie bureaucratic calm. A broad context window can turn a messy organization into a voice that sounds settled.

The DevTools example is valuable because it starts from constraint rather than appetite. It does not ask the model to become wise by swallowing the whole trace. It designs a way for the assistant to ask for the next piece of evidence in a known form. A useful assistant needs paths, not just piles.

That lesson travels beyond debugging. An agent working in a repository should not begin by absorbing every issue, pull request, architectural debate, and ancient TODO with equal enthusiasm. It should start from the task, the files likely to matter, the tests that protect the behavior, and the recent decisions near the change. It should have ways to widen the view when the evidence demands it, and ways to say when the available context is not enough.

A workplace agent preparing a weekly brief should not behave like a vacuum cleaner with a calendar integration. It should know the difference between commitments, drafts, customer evidence, policy, and someone's temporary attempt to think in public. Some context should summarize. Some should cite. Some should expire. Some should remain off-limits unless the person explicitly opens the door.

Microsoft's 2026 Work Trend Index is useful here because it turns the spotlight away from individual prompting and toward organizational conditions: shared practices, manager support, workflow redesign, and structured use of agents. The useful conclusion is not that everybody needs a bigger model. It is that AI value depends on the design of the surrounding work.

Blinders are part of that design.

That word can sound punitive, as if the agent is being made dumber for safety's sake. I mean something closer to craft. A microscope is useful because it excludes most of the room. A good function accepts a narrow contract and refuses the rest.

Agents will need the same dignity of refusal. They will need interfaces that make their field of view visible: what they looked at, what they did not look at, what they can request next, and what kind of context was deliberately withheld. They will need task scopes small enough to inspect and broad enough to be honest.

This is not an argument for timid tools. It is an argument for tools with posture. The more powerful the agent becomes, the more important its attention design becomes. A weak assistant can be annoying when it sees too little. A strong assistant can be dangerous when it sees too much and understands too softly.

The next stage of workplace AI may arrive with more connectors, wider context windows, richer graphs, and longer-running tasks. Some of that will help. But the better agents will not be the ones that can see everything all the time. They will be the ones that know how to enter a room, ask what kind of room it is, look only where the work justifies looking, and leave the rest of the human weather alone.

The future of context is not maximum context. It is situated attention. The useful agent has blinders, and the humane organization knows how to fit them.