No. 020 Wednesday · May 27, 2026 ≈ 5 min read

Agent consoles need attention state.

The week’s practical signal is not another blank prompt box. It is the arrival of work surfaces that show when agents need review, permission, alert context, or a browser beside the terminal.

Parallel sessions Permission lanes Alert review
Abstract agent operations console with split terminal panes, browser preview, notification rings, permission chips, and alert rows.
Agent ops plate: terminal, browser, permission, alert
Today's Art Direction

Agent Ops Console / Attention State System

A command-center grammar for agent work that keeps asking for human judgment.

Ops consoles are built for interruption without chaos: the most important state gets a visible ring, permission is separated from telemetry, and the operator can jump from a summary to the live surface. For agentic web design, the reusable move is to make attention a first-class component instead of hiding it inside chat history.

Attention RingPermission LaneSplit PaneAlert RowRun QueueBrowser DockStatus Light
§01Tooling
Tooling

The terminal is becoming an operations surface.

cmux packages the multi-agent desk.

cmux is a Ghostty-based macOS terminal for coding agents, with vertical tabs, split panes, notification rings, and an in-app browser. The design signal is blunt: once several agents run in parallel, the terminal needs workspace metadata, browser state, and attention markers in the same frame.

Vercel moves investigation into the CLI.

Vercel’s alerts command can now show anomaly details in the terminal, and the AI option brings investigation results alongside each alert. That makes production debugging feel less like checking a dashboard and more like handing an agent an already-labeled incident row.

The interface job is no longer “make a place to type.” It is “make a place where parallel work can ask for the right kind of attention.”

§02Technique
Technique

Design the attention state before the action.

Separate waiting from working.

cmux’s notification rings are useful because the visible state changes only when an agent needs you. Borrow that restraint: not every run needs a glow, but every blocked run needs a clear target.

Name the permission lane.

Claude Code’s permission modes distinguish planning, automatic file edits, auto decisions, and bypassed checks. Treat those as UI lanes with different risk, not as hidden command settings.

Keep the browser nearby.

For web work, a terminal-only loop misses the thing users inspect. A split browser preview turns an agent action into a state the designer can judge immediately.

§03Workflow
Workflow

Let the console carry the handoff.

Use a two-column agent review layout.

Put the running transcript, diff, or task log on the left, and the live artifact on the right. The operator should be able to answer three questions without scrolling: what changed, what needs permission, and what surface proves the result.

Bring design context into the loop.

Figma’s Dev Mode index now foregrounds workflows like Expanding the canvas with Figma MCP, where real product states return to the canvas for review. The practical pattern is the same as the ops console: bring evidence back into the place where decisions are made.

The next agent interface is less a chat window than a supervised workbench with visible wait states.

§04Prompt Lab
Prompt Lab

Ask for an attention-state plan.

Use this before a coding or design agent starts a multi-step task that may need review, browser inspection, or permission changes.

Before implementation, define the attention states for this task.

For each state, list:
- what is happening
- what evidence should be visible
- what permission is required
- what browser or artifact view proves the result
- what signal means the agent needs human review

Then start only with the first state.

The prompt is short because the interface should do the heavy lifting. The goal is to make the agent declare where it expects review before it begins changing the work.

§05Field Note
Field Note

Attention is a design material.

Hacker News also had a fresh daily-driver guide to Claude Code on the front page, which is a good cultural marker: agent workflows are moving from novelty into operating habits. The next design problem is not whether agents can work; it is whether the surrounding interface makes their state legible enough to supervise.

The strongest agent products will make interruption feel designed: visible, scoped, recoverable, and tied to the artifact being changed.

§06Sources
Sources

Primary links for the day.

Signal scan
Digg AI rankings — live AI-Twitter and GitHub-stars pulse check.
Tooling
cmux on GitHub — Ghostty-based terminal with tabs, browser panes, and notification rings for coding agents.
Ops
Pull anomaly alert details using the Vercel CLI — Vercel changelog, May 21, 2026.
Permissions
Choose a permission mode — Claude Code documentation on plan, accept-edit, auto, and bypass modes.
Design loop
Figma Dev Mode blog index — includes the April 30, 2026 workflow lab on Figma MCP and canvas review.
HN signal
Hacker News front page — Claude Code daily-driver guide appeared on the front page during the research pass.
Practice
Beyond the Prompt: Claude Code — Arpan Patel’s May 26, 2026 guide to Claude Code habits, skills, subagents, and MCPs.

A field experiment from the team behind Beaver Builder.