Published on

Claude Code's New Eyes β€” Agent View and /goal Let AI Run to the Finish Line

Tell a runner to give their best effort without knowing where the finish line is. They'll run hard β€” but they won't know when to stop. That has been the quiet limitation of AI coding tools up until now.

On May 11, 2026, Anthropic shipped Claude Code v2.1.139. Two features lead the update: Agent View and the /goal command. On the surface, they look like a UI improvement and a new command. In practice, they change the operating philosophy of AI coding tools.

Agent View: Every Session in One Place

Claude Code Agent View Dashboard

A familiar problem emerges when you work with AI coding tools at scale. Feature A is being built in session one, a bug is being fixed in session two, docs are being written in session three β€” and keeping track of what's happening where starts to feel like a job in itself.

Agent View addresses this head-on.

Type claude agents in any terminal or press the left arrow key. Every running Claude Code session appears in a single list. Each row surfaces four signals at a glance:

  • Session ID β€” which session this is
  • Waiting status β€” whether it's blocked, waiting for your input
  • Last response β€” what the AI said most recently
  • Timestamp β€” when the last interaction happened

Think of it as an air traffic control tower for your AI agents. You can monitor every active session, respond to the ones waiting on you, and see the overall picture without switching context repeatedly.

Agent View is currently in Research Preview and is available to users on Pro, Max, Team, Enterprise, and Claude API plans.

/goal: AI That Knows Where the Finish Line Is

The most philosophically interesting feature in this update is the /goal command.

Today's typical Claude Code workflow works like this: you give an instruction, the AI executes one step and stops, you give the next instruction. Claude is an excellent executor β€” but it doesn't independently track whether the ultimate objective has been achieved.

/goal reverses this pattern.

/goal all tests must pass and the CI pipeline must be green

With that single declaration, Claude Code keeps iterating automatically until the condition is met. If tests fail, it analyzes why and modifies the code. It runs the tests again. If they still fail, it tries a different approach. It loops until the condition is satisfied.

While the work runs, an overlay panel shows live status:

MetricMeaning
Elapsed timeHow long since the task started
Turn countHow many steps the AI has executed
Tokens usedRunning token consumption for cost tracking

Users can also set constraints upfront β€” a maximum time limit, maximum turn count, or token budget β€” to prevent runaway loops. /goal works in interactive mode, with the -p flag, and in Remote Control mode.

Real-World Scenarios: Where /goal Shines

Here are a few situations where /goal changes things meaningfully:

Scenario 1: Legacy Code Migration

/goal all CommonJS require() statements are replaced with ES Module 
      imports and npm run build completes without errors

Instead of manually converting hundreds of files, you declare the end state and let Claude find the path.

Scenario 2: Test Coverage

/goal test coverage for the new feature module is at or above 80%

Claude decides what tests are needed, writes them, and verifies coverage β€” looping until the threshold is met.

Scenario 3: Accessibility

/goal the Lighthouse accessibility score is 95 or higher

Check score β†’ fix issues β†’ recheck. Claude does this loop automatically.

An EdTech CEO's Take: Goal-Oriented Learning, Mirrored in AI

Education theory has long debated the difference between process-oriented and outcome-oriented learning. The /goal command lands firmly on the outcome side. Instead of "do these steps," it says "this state must be true." Claude finds its own path to get there.

The parallel to education is worth noting. Telling a student to "read chapter 1" is process-oriented. Saying "you should be able to explain this concept in your own words" is goal-oriented. Both matter. The /goal command is a reminder that framing the target clearly is often more powerful than prescribing every step.

The caution, though, is real. When AI drives toward a goal across multiple iterations, a developer who doesn't fully understand the path may end up with code they can't explain or maintain. The vibe-coding generation needs to ask: am I learning alongside this, or just collecting outputs?

Tips for Getting Started

  1. Open Agent View: Run claude agents or press the left arrow in any session. Understand which sessions are running vs. waiting on you before doing anything else.

  2. Write verifiable goals: "Work well" is too vague. "All unit tests pass" is testable. The more precisely you define the finish line, the more effectively Claude reaches it.

  3. Set constraints first: On your first few runs, set a low turn or token limit. This lets you catch Claude heading in the wrong direction before it goes too far.

  4. Break large goals into stages: A series of verifiable intermediate goals is more effective than one massive goal. Run /goal iteratively.

  5. Combine with Agent View: Run multiple goals in parallel across different sessions, then monitor all of them from Agent View. That's the full multi-agent workflow.

Sources

Discuss on Twitter β€’ View on GitHub
Claude Code's New Eyes β€” Agent View and /goal Let AI Run to the Finish Line | MINSSAM.COM