Claude Code gets a small control panel for agent work

This sounds like a dry release note, but it is useful: Claude Code can now list live agent sessions as JSON. If you run several AI jobs side by side, Claude starts to look less like a black box and more like a work queue with lights: running, waiting, needs a human, done.

Source: Claude Code changelog 2.1.145

What is new in Claude Code 2.1.145

A coding agent is an AI assistant that can read code, edit files, run commands, and carry a work goal across several steps. In Claude Code 2.1.145, published on May 19, 2026, the most practical addition is claude agents --json: a machine-readable way to list active Claude sessions.

That makes small tools around agent work possible:

• Status lines that show which Claude jobs are waiting for input
• Custom terminal pickers or tmux restoration
• Simple internal views of agent, repo, PR, and latest event
• Clearer logs when a subagent job belongs to a parent job

The same release also adds agent_id and parent_agent_id to OpenTelemetry spans, lets status line JSON include GitHub repo and PR information, and includes background_tasks and session_crons in Stop and SubagentStop hooks. OpenTelemetry, often shortened to OTel, is a standard for sending metrics, logs, and traces to monitoring tools.

Source: GitHub release v2.1.145

Why this matters in practical work

If a consultant, product owner, or small development team only runs one Claude chat at a time, this may feel like too much. Once Claude Code has several jobs, for example one writing tests, one investigating a bug, and one preparing a PR, you need a simple picture of the state of work.

The question is not only "can Claude do the job?". It becomes: which job is waiting for me, which repo does it touch, which PR belongs to the change, and what do I need to read before approving anything?

This is where the JSON list, status line, hooks, and OTel start to fit together. They let you build a small agent panel without buying a whole new work-management tool first. For Hammer readers, this is classic Tool Forge work: make AI work visible, measurable, and stoppable before it touches real systems.

Source: Customize your status line

Integrate safely without making AI useless

This kind of integration does not need to start as a huge security project. Start with a few concrete rules.

Prefer logging metadata first: session, repo, PR, cost, model, status, and which tools were used. Do not store raw prompts, customer data, or file contents in logs just because the system allows it. If you send OTel data to a collector, use environment variables or a secret manager for tokens, keep permissions narrow, and add redaction where sensitive paths or customer names could leak.

The Claude Code docs also describe sensitive flags such as OTEL_LOG_USER_PROMPTS, OTEL_LOG_TOOL_DETAILS, and OTEL_LOG_RAW_API_BODIES. They are powerful. Leave them off until you know why you need them, where the data goes, and who can read it.

Source: Monitoring Claude Code usage

Try this prompt this week

Use it in Claude Code after updating to the latest version. Run it in a repo where you are allowed to experiment with settings, hooks, and the status line. Do not connect real secrets in the first pass; use environment variables, scoped tokens, and a local test collector if you want to try OTel.

You are my Claude Code ops designer. Help me build a simple control panel for agent work in this repo.

Goal: I want to see which Claude Code sessions are running, which ones are waiting for input, which repo/PR they belong to, and which jobs need human review.

Use these signals:
- claude agents --json
- status line JSON with repo and PR information
- Stop and SubagentStop hooks with background_tasks and session_crons
- OpenTelemetry for metadata, metrics, and traces

Before suggesting code, do this:
1. Describe which fields we should show in a minimal agent overview.
2. Suggest a status line that helps without getting in the way.
3. Suggest which hook events we should log and which ones should block or require approval.
4. Say exactly what we should not log in version one, especially prompts, customer data, file contents, and secrets.
5. Suggest how tokens should be handled with environment variables or a secret manager and least-privilege permissions.
6. Give me a 30-minute setup plan with commands, files, and test steps.
7. End with a human review checklist before this is used in a real project.

A good result looks like this:

• You can see which Claude session is waiting without opening five terminals.
• Every job has a repo, branch or PR, owner, and latest event.
• Logs contain metadata, not raw customer text.
• Hooks stop or pause risky actions until a human approves them.
• A new colleague can understand the setup in ten minutes.

What we are watching next

On the same day Claude Code got better agent tracing, Anthropic also announced a broader KPMG alliance where Claude will be used by more than 276,000 employees and embedded into KPMG Digital Gateway. Most Hammer readers do not need to think at that scale. But the direction is clear: when Claude moves from chat into the work environment, visibility, responsibility, and human review become part of the tool itself.

The next practical step is to choose one workflow where Claude Code already helps you and put a thin control panel around it. Not to slow the work down. To say yes faster when the evidence is actually there.

Source: KPMG integrates Claude across its core business and workforce of more than 276,000