Continue a coding task by email
Start one clear task from your phone, then keep the follow-up in the same email thread. The agent runs against the project on your own computer.
Watch the existing product walkthrough
The 2 minute 12 second film follows an emailed task to the runner on your own computer and the replies in the same thread. It was published on 16 October 2026. Use the current instructions below for setup and approval behavior.
Watch the product walkthrough on YouTube. The written steps below cover starting a task, continuing the session and reviewing a blocked action.
Start after the runner is ready
Install and connect the runner using Getting started. The computer with your project must stay awake and online. Wait for the ready message before sending the first task: the initial successful intake skips old Sent mail, so old messages are not treated as new work.
@gent reads authenticated Sent mail from your account. Email yourself at the address with +agents added and use the agent's supported subject form. This lets you send work from another device without exposing a new public endpoint on your development computer.
Write a task you can review
Name the part of the project to change, the behavior you want and the check that would show it works. Include boundaries that matter, such as keeping the existing design or preparing a change for review before release. Avoid a vague instruction that leaves the agent guessing which project or outcome you meant.
To: you+agents@gmail.com Subject: Codex: improve the signup form In the configured project, make the email error explain how to fix it. Keep the existing components and spacing. Check the valid and invalid states and report the changed files.
Use the project aliases and configuration to select the right working folder. The runner executes on that computer with the configured agent and permissions; email is the control surface, not a separate hosted development machine.
In the configured project, make the email error explain how to fix it. Keep the existing components and spacing. Check the valid and invalid states and report the changed files.
Queued: codex / main / agent default
Started. Your agent is working on this task.
- Action
npm test -- signup- Folder
~/code/web- Reason
- Run the signup form tests.
- Task
3f9c2a7e
approveone action, not the sessionCompletedThe email error now says what to change. I kept the existing components and ran the signup form tests. They pass.
src/components/SignupForm.tsxsrc/lib/validate-email.ts
Use the existing error component and keep the input focused after rejection.
- 1The task. Name the project area, the behavior and the check that shows it works.
- 2Queued and Started. Each update arrives as a reply in the task's thread.
- 3One approval. Read the action, folder and reason. Reply to that email with exactly approve or deny.
- 4The result. Read which files changed and which checks the agent actually ran.
- 5The follow-up. A reply in the same thread continues the session. Name what should change.
Reply to keep the context
You receive Queued, Started and the result in the same thread. Reply to continue that session with a correction, a narrower request or the next step. A useful follow-up names the part of the result that needs changing and what should replace it.
Make the next instruction concrete.
"Use the existing error component and keep the input focused after rejection" gives the agent an observable result to implement. A general reaction gives it less evidence about the intended fix.
Approve one blocked action
When email approvals are enabled, an action needing permission pauses the task. Read the complete action, working folder and reason in the approval email. Reply directly to that email with exactly approve or deny. A reply in the task thread uses the token from the approval email.
An approval covers one action. It does not grant every command in the session. A plain "yes", a reused token or extra quoted text does not count. Approval emails documents the timeout, supported agent versions and refusal cases.
Check the result before the next request
Read which files changed and which checks the agent actually ran. If a check failed, continue from that specific failure. If the computer went offline or the runner did not process the message, use Troubleshooting rather than repeatedly mailing the same instruction.