How task execution works

This page walks through what the Coding Agent does between you clicking Approve on a plan and Andy posting a pull request. It's useful when something looks stuck or unexpected — knowing the steps helps you tell whether to wait, intervene, or stop.

For the full picture of how a task moves from request to merged PR, see Task lifecycle.

What happens when you approve a plan

The task moves to Implementing.
Andy posts a "Working on it…" message in the Slack thread with a Stop button.
The Coding Agent clones your repo to an isolated environment and starts a fresh branch for the change.
Tests run if the plan asked for them — the agent attempts your project's standard test command and includes the result in the walkthrough.
Andy updates the thread with brief progress notes as it works. Most tasks finish in a few minutes; complex tasks can take longer.

If the agent gets stuck or needs clarification, it pauses and sends a message in the thread — reply normally to unblock it.

The walkthrough step

When the Coding Agent finishes coding, it doesn't immediately open a PR. Instead, Andy posts a walkthrough summary in the thread:

Implementation complete

Task: [task title]
Branch: [View branch]

[ 📄 View implementation details ]

[ 🚀 Create pull request ]   [ ✏️ Make changes ]   [ 🗑️ Discard and exit ]

The walkthrough on the task details page shows what changed — files touched, before/after snippets, test output, and any caveats Andy flagged. This is your chance to ask for revisions before a PR exists, which is cheaper than asking for changes after.

Three options:

Create pull request (or Create merge request for GitLab) — Andy opens the PR/MR. The task moves to PR open.
Make changes — reply in the thread or in the web app with what you want adjusted. The Coding Agent pushes follow-up commits to the same branch and updates the walkthrough.
Discard and exit — throw away the branch. The task moves to Abandoned and the work is gone.

Stopping mid-implementation

The Stop button is available throughout planning and implementation. Clicking it:

Aborts the current Coding Agent run
Posts ✅ [Your name] stopped this task in the thread
Discards any work that hasn't been committed to the branch yet
Leaves the task in its current status — you can ask Andy to retry, or mark it as done to close it out

You'd use Stop if you realize the plan was wrong, the task isn't worth finishing, or the agent has clearly gone off course.

When the Coding Agent fails

If the Coding Agent hits an unrecoverable error — for example, it can't make the change cleanly, or your test suite is broken in a way it can't fix — Andy reports the failure in the thread with the error details. Common next steps:

Reply with more context to give the agent a hint and let it try again
Stop the task and start over with a more specific plan
Mark as done if the partial work is enough or the task is no longer relevant

The Coding Agent does not retry automatically — failures surface fast, on purpose, so you can decide what to do.

Mark as done

For tasks that ended up resolved outside the normal flow (e.g., a teammate did the work manually, or it turned out not to be needed), you can mark them done from the three-dot menu on the task details page. This sets the status to Done without needing a PR.

What's next

Reviewing Andy's work →