Agents need a record of scope.

Codex turns context capture into a gesture

OpenAI's May 21 Codex update adds Appshots on macOS, Goal mode across the app, extension, and CLI, and browser annotation improvements. The useful part for builders is mundane but large: instead of describing the current window, a user can attach the app state and available text directly to the work thread.

That pushes agent UI away from blank prompt boxes and toward working surfaces. A design review, doc edit, or failing browser state can become part of the instruction record, which means the interface has to show what context was supplied and where the agent's confidence is coming from.

Operations data moves into the terminal

Vercel's new alerts command brings anomaly details and AI investigation results into the CLI. This is not only a convenience feature; it gives coding agents a structured path from alert to action without scraping dashboards or asking a human to paste screenshots.

For design engineers, the pattern matters more than the command. Agent-ready tools increasingly expose a compact, authenticated, text-native version of the product state, and that state becomes part of how a workflow is designed.

• SDK / tools

  Vercel gives chat agents role-based tool presets.

  Reader, messenger, and moderator presets make tool access feel like an interface choice, not a hidden model capability.

  Steal this: name the role before exposing the tool.
• Safety / scope

  Overeager Coding Agents measures out-of-scope edits.

  The benchmark is useful because it treats overreach as its own failure mode, separate from whether the model can solve the task.

  Steal this: make consent language explicit.
• Workflow / review

  Every's automation report keeps humans in the loop.

  The interesting claim is not that work disappears. It is that human work shifts toward framing, review, taste, and differentiation.

  Steal this: design the review surface first.

• 1 / context

  Agents are getting better source material.

  Appshots, CLI alert records, and structured SDK tools all reduce the need for humans to narrate state by hand.

  Design question: what did the agent actually see?
• 2 / boundary

  More context raises the cost of vague permission.

  When an agent can inspect more surfaces, the product needs clearer limits around writes, messages, file edits, and follow-up actions.

  Design question: what is outside the task?
• 3 / evidence

  Agent work needs receipts.

  The useful artifact is a compact record of sources, decisions, commands, and approval points that a human can scan later.

  Design question: where is the scope ledger?

Before editing, give me a scope record:

1. Goal
2. Context you will use
3. Files or surfaces you may touch
4. Things you will leave alone
5. When you will stop and ask

Then make the smallest useful change.

• Bottom line

  Agents are gaining eyes, tools, and reach.

  The design opportunity is the record that keeps those capabilities bounded, reviewable, and easy to hand back to a human.