# How do I see what my Claude Code and Codex agents are working on?

_See what coding agents are doing by tracking ownership, status, decisions and proof in one shared task record outside their chat sessions._

By Saad Ahmed, 2026-10-20

To see what Claude Code and Codex are working on, make every agent claim work and report progress to one shared task record. That record should show the current owner, status, last meaningful update, blocked decisions and links to the work itself.

Agent chat windows are execution surfaces, not a reliable portfolio view. If each session keeps its own list, you have to open several transcripts and infer whether the work is current.

## Track outcomes, not activity theater

A useful view answers five questions for every task:

- What outcome is the agent responsible for?
- Which agent owns the next action?
- Is the task ready, in progress, waiting for a person or finished?
- What changed since the last update?
- What branch, pull request, document or test proves the claim?

Avoid dashboards that show only tokens, running processes or elapsed time. Those signals can help diagnose a session, but they do not tell you whether the intended result exists.

## A repository-only setup

For a small project, keep an `AGENT_WORK.md` file with one row per task. Require each agent to add its name, branch, status and latest evidence before and after working. Pair that with your Git host so you can inspect the branch and pull request.

This is inexpensive and transparent. Its weakness is coordination: two sessions can update the file at once, stale rows need manual cleanup and a question for you can disappear inside a transcript.

## Use one shared board for several agents

When multiple agents work across repositories or sessions, use a board they can all read and update. Each agent should claim a task before changing files, keep the status current, and attach evidence when it stops.

ecco gives people and MCP-connected agents the same board. It does not run the agents; it holds the shared state around their work. The overview shows what is active, while the inbox separates items that need a human decision from work agents can continue themselves.

[Connect an agent to ecco](/blog/connect-an-agent), then instruct Claude Code and Codex to read the board before selecting work. For the difference between a shared board and assistant-local task features, see [ecco versus Claude and ChatGPT](/blog/ecco-vs-claude-and-chatgpt).

## Set a reporting contract

Give every agent the same short operating rule:

```text
Before work, claim one ready task and record your branch.
During work, report decisions or blockers on that task.
When stopping, attach the result and checks, then set the honest status.
Never mark done without evidence.
```

The contract matters more than how often the board refreshes. A current, evidence-backed update is more useful than a live stream of unstructured actions.

## Frequently asked questions

### Can I monitor Claude Code and Codex in one place?

Yes, if both agents read and update the same external task system. Use a repository file for a small sequential workflow or an MCP-connected board when agents work in parallel.

### Should I track every command an agent runs?

Usually no. Track the outcome, ownership, meaningful status changes, decisions and completion evidence. Use command logs only when they help investigate a failure or verify a sensitive change.

### How often should an agent update its status?

At minimum, update when claiming work, reaching a meaningful checkpoint, requesting a decision, becoming blocked and finishing. Long tasks may need additional concise progress notes.

### Does a pull request prove the task is complete?

A pull request proves that a proposed change exists. Completion may also require passing checks, review, merge, deployment or a live verification, depending on the task's definition of done.