How do you give an AI agent a long job without losing track?

Why should a long job be a checklist the agent ticks?

A long job fails in the middle more often than at the start. If the only record is a paragraph of intent, nobody can tell which part ran, which part was skipped, and which part was invented to look complete. A checklist is a contract: each line is a unit of work small enough to finish or fail on its own.

Write the list before the first tool call. Keep items observable. Prefer verbs that leave an artefact you can open: a draft, a sent message, a deployed preview, a trade idea written as rules. Avoid items such as think harder or polish everything. The agent should tick only after the artefact exists, not after it has described what it meant to do. A notepad or a shared document is enough. A simpler to-do list in a notes app is the better choice when the work is one person and one sitting.

• One outcome per line, written so a stranger could check it.
• Tick only when the artefact is present, not when the plan sounds finished.
• Leave blocked items visible instead of deleting them to tidy the list.

Why does a list you can see beat a summary at the end?

A closing summary is a story. It can reorder events, merge two attempts into one, and hide the step that never ran. A live list is a ledger. You can interrupt at any moment and still know the last honest state. That matters when the work spans email, a build, a calendar hold and a later review: the thread is the ticks, not the chat scroll.

Summaries still help after the fact, as a letter to your future self. They should point at the list, not replace it. If a tool only offers a final recap, keep the checklist outside it. GROX keeps memory between sessions and can run unattended work, but the same rule applies: the human-readable ticks are the source of truth, not a fluent wrap-up.

Open

Work not started or not evidenced yet.

Ticked

The named artefact exists and was opened or read back.

Blocked

The step cannot finish; the reason is written on the line.

Closed without delivery

You accepted a partial result and recorded what was dropped.

How do you stop partway and carry on later from the ticks?

Stopping is not a failure if the list still tells the truth. Pause when an item needs a human decision, a login you do not want to automate, or a night’s sleep. Write one line under the current item: what was tried, what is waiting, what must not be repeated. Then leave. Do not ask the agent to remember in prose alone.

When you return, start from the first open tick, not from a new brief. Restate only the constraints that changed. If a sandbox or a draft was destroyed, the list should say rebuild from this file rather than start the product again. Autopilot and overnight runs are useful only when each tick still has a clear stop: size limits, a human ask before spend, or a hard pause in a sentence. If the job is a single document you will finish yourself, a simpler editor with headings is the better choice than an agent at all.

What should done mean when one part could not be finished?

Done is not a mood. It is every line either ticked or closed with a reason you accept. If one part could not finish, name the missing piece on the list, say what was delivered instead, and say whether a follow-up is scheduled. Do not bury the gap in a cheerful last paragraph.

A useful close looks like this: the pages that run are listed; the page that does not is named; one automatic retry was already used if the process allows it; further retries wait for a cause, not another patch. Shared work should make the unfinished item public to anyone who inherits the job. That is how you keep the thread without pretending the whole job landed.