Where agents fit first

The first question after deciding to run work through agents is which work to hand over. The answer is not "the most painful work" or "the most impressive demo" — it is the work where an agent can succeed on its own terms and you can tell that it did. This page is a filter for choosing that work, so your first weeks with agents build trust instead of cleanup.

Start with work that repeats, has a clear bar, and stays reviewable

Three properties make a job a good first candidate. It is recurring, so the effort of defining it pays back every cycle instead of once. It has clear success criteria, so both you and the agent know what done looks like — a draft exists, a report covers the agreed sections, the research answers the question that was asked. And it is internally reviewable: the output lands in front of you before it goes anywhere that matters, so a weak result costs you a review, not a customer.

In practice that filter points at a familiar set of jobs. Content drafts — posts, newsletters, changelog entries — repeat on a schedule and are judged the moment you read them. Research and competitive summaries have a defined question and a written answer you can check against sources. Reporting pulls the same numbers into the same shape every week. Outreach drafts are written, reviewed, and only then sent. All of these produce artifacts you inspect at your desk, which is exactly where you want an agent's early work to land.

A worked example: handing over the weekly changelog

To see the filter in action, take a job most product companies carry: the weekly changelog. It passes all three tests. It repeats every week, so the setup pays back every Friday. Its bar is clear: every shipped change covered, written in your voice, every link pointing at something real. And it is reviewable — the draft sits in front of you before anyone outside the company sees it.

Handing it over is a matter of giving the job its shape. The workflow gathers the week's shipped changes, drafts the entries against your voice document, runs a verifier step that checks each entry links to a real change, and stops at an approval step before anything publishes. The first Friday, the draft lands in your inbox and you rewrite two entries that read like engineering notes instead of customer language — and that correction goes into the voice document, not into a chat that forgets it. The second Friday needs one small edit. By the third or fourth, you are approving unchanged, and an hour of your Friday has become a two-minute decision — with the run's step-by-step history there if an entry ever looks off.

The same shape transfers to its neighbors. Swap the gathering step for prospect research and you have outreach drafts. Swap it for last week's numbers and you have the weekly report. Once you have handed one job over end to end, the second is mostly recognizing the pattern.

Keep judgment-heavy, externally risky work supervised

The mirror image also matters: work that is high-judgment or lands directly on the outside world is not where agents start unattended. Pricing decisions, replies in a tense customer thread, anything legal or financial, a message sent under your name to someone who matters — an agent can help draft all of it, but the send stays behind your approval. This is not a permanent restriction. It is the honest starting position. Agents in Task Machine begin supervised precisely so that this class of work flows through your inbox before anything irreversible happens, and you loosen that boundary deliberately as results earn it.

The distinction is not "easy work versus hard work". Agents handle genuinely hard research and writing well. The distinction is how expensive a miss is and how quickly you can catch one — internal and reviewable first, external and consequential later.

Playbooks are the starting shape

You do not have to design your first agent job from a blank page. Playbooks are ready-made bundles — the agents, workflows, and documents for one recurring job — that you install and adapt rather than assemble. Installing a playbook for, say, weekly content gives you a working shape on day one: an agent with instructions, a workflow with the review step already where it belongs, and documents that tell the agent how your company talks. Your first cycles are then spent correcting output in the task timeline and tightening the documents, which is far faster than inventing the process while also evaluating the agent. Playbooks covers how they are put together.

Once the first job is running, the question changes from "what should agents do" to "how do I stay in charge without reviewing everything myself". That is a real tension, and the product is built around resolving it — Control without becoming the bottleneck is the next page.