How to Delegate a Multi-Step Task to an AI Agent Safely

9 min read

Make this article actionable

Send the article context into Vife Agent and turn it into a plan, checklist, or draft you can keep working on.

Open in Agent

Most people hand an agent a goal and then find out, forty minutes later, that it has also sent two emails, renamed a folder, and bought something.

The failure is rarely the model being stupid. It is that "prepare next week's launch" is not a task; it is a wish with a deadline attached. An agent that can browse, read your files, and write documents will do something with a wish — just not the something you had in mind, and not in an order you approved.

This guide is the brief I now write before delegating anything longer than a single lookup. It takes about fifteen minutes, it is tool-agnostic, and it works the same whether the agent runs inside a chat product, a coding tool, or a workspace like Vife where an agent task can use a browser, read the files you attach, and produce a document, a deck, an image set or a short clip.

Why this is the week to fix your delegation habit

Three launches in nine days all push in the same direction: the assistant is being sold as something that finishes work, not something that answers a question.

  • November 17: OpenAI made GPT-6.2 Astra and Sol available in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, with API access as gpt-6.2-astra and gpt-6.2-sol and published API prices of $8 input / $40 output per million tokens for Astra, and $2 / $10 for Sol (OpenAI Developer Community, November 17, 2026).
  • November 9: Anthropic introduced Claude Fable 5.2 and Claude Mythos 5.2, describing Fable 5.2 as usable either as an "Advisor" consulted by smaller models or as an "Orchestrator" that delegates simpler work to them. The same page lists $8 input / $40 output per million tokens for standard mode and $16 / $80 for fast mode (Anthropic, November 9, 2026).
  • November 10: Google said Gemini Intelligence is available on Pixel 10 devices and showed multi-step tasks such as getting groceries, booking a ride and ordering takeout, in Japan, the UK and India (Google, November 10, 2026).

The same Anthropic page also shows the gap you have to plan for. It reports Fable 5.2 at 51.4% on AutomationBench and 41.3% on the Remote Labor Index — vendor-run figures, not independent audits, and nowhere near 100%. Put those two numbers next to the launch copy and the practical conclusion is neither "agents are useless" nor "agents are solved." It is that a multi-step task will sometimes come back half-finished, and you need to be able to see which half.

That is a briefing problem more than a model problem. (That reading is mine, not the vendors'.)

Mid-read shortcut

Turn the useful parts into next steps

Vife Agent can convert this guide into a prioritized workflow with tasks, risks, and reusable prompts.

Create a brief

The difference between a prompt and a delegation brief

A prompt says what you want. A delegation brief says what you want, what the agent may touch, when it must stop, and how you will check the result.

Six fields. If you can fill all six in a page, you can delegate almost anything that fits inside a working day.

1. One outcome, one artifact

Write the finish line as a noun you can open, not a verb you can argue about.

  • Weak: "Research competitors and help with positioning."
  • Strong: "A six-section markdown brief comparing three competitors on pricing, onboarding steps and stated integrations, with a source link in every section."

The artifact decides when the task is over. It also decides what "done" looks like when you review it, which matters more than any prompt wording. If you cannot name the file, you are not ready to delegate.

2. Split read steps from write steps

Every multi-step task is a mix of looking and changing. Separate them before the agent starts, because the two carry completely different risk.

  • Read steps: open pages, read the attached PDFs, extract a table, compare two versions. Getting one wrong costs you a correction.
  • Write steps: send, publish, upload, book, buy, delete, edit a shared file. Getting one wrong costs you an apology.

Write the brief so the agent finishes all read steps and proposes the write steps as a list. In practice that means a line like: "Do not send, publish, upload, book or delete anything. At the end, list the exact actions you would take, with the recipient or URL for each."

You will be surprised how often the proposed list contains something you did not want — and how cheap that discovery is compared to the alternative.

3. Draw the stop line explicitly

Agents fail expensively in three places: money, messages, and accounts. Name all three.

Stop and ask before: any spend, any message to a person outside this task, any change to an account setting, any deletion, and anything involving credentials. If a page asks you to sign in, stop and report the URL instead.

That last clause matters because a login wall is where an agent starts improvising. A stop line converts a silent failure into a message you can answer.

4. Give it a source of truth and a citation rule

An agent with no allowed sources will happily reconstruct a plausible version of your company from memory. Attach the real material and say which material wins.

A working rule: "Use only the attached files and the URLs listed below. If a fact is not in them, write 'not found' rather than filling the gap. Put the source link at the end of the sentence it supports."

This is the single highest-leverage line in the brief. It turns the output from an essay into something you can audit in a minute, and it makes the "not found" cells visible instead of invisible.

5. Dry-run on one slice

Never let the first run be the whole job. Give the agent a tenth of the work: one competitor instead of three, the first four pages instead of forty, one product photo instead of the set.

The dry run answers three questions cheaply:

  1. Did it understand the artifact?
  2. Did it respect the stop line?
  3. Is the quality ceiling high enough to be worth the remaining nine-tenths?

If the answer to any of those is no, you have lost ten minutes rather than an afternoon. In Vife specifically, this is also the moment to check that the files you attached are the ones the agent actually read — the task view shows which inputs it opened, and that is usually where a wrong-source problem first becomes visible.

6. Check against an acceptance test, not a vibe

Write the acceptance test before you read the output, so the output cannot redefine success.

For a research brief: "Every claim has a link; every link opens; the three competitors each have a price and an onboarding step; no section is longer than 150 words."

For a document: "The decision is on page one; the numbers match the attached sheet; nothing is marked 'TBD' that I did not mark myself."

For an image set: "Same crop, same light direction, product color unchanged across all six."

The test should be mechanical enough that a colleague could run it without you. If it needs your taste, it is a review step, not an acceptance test — and the brief should say so.

A worked example: the Tuesday launch brief

Here is the whole thing as one paste-able block. It took eleven minutes to write and has saved more than one ruined afternoon.

text
Outcome: A one-page launch checklist as markdown, covering the six assets we need for the March feature launch, each with an owner and a due date taken from the attached project file. Sources: attached "launch-plan.md" and "asset-list.csv" only. Nothing from memory, nothing from the web. If a field is missing, write "missing". Read steps: read both files, reconcile the two asset lists, flag every disagreement between them in a short section at the end. Write steps: none. Do not send, upload, publish, book or delete anything. Stop line: stop and ask if the files disagree on a launch date, or if any task requires signing in somewhere. Dry run: do the first two rows only, then stop and show me the format. Done when: six rows, no invented dates, disagreements listed separately, output under 400 words.

Note what is absent: no tone instructions, no persona, no "think step by step." Those are prompt decoration. The brief is mostly constraints, because constraints are what make a multi-step run checkable.

Checklist before you press go

  • The outcome is a named artifact, not a verb.
  • Read steps and write steps are listed separately.
  • The stop line names spend, messages, accounts, deletion and credentials.
  • Allowed sources are attached or linked, and the agent is told to write "not found" instead of guessing.
  • A dry run on one slice is defined.
  • The acceptance test is written down before the run.
  • The task has a visible output location you will actually open.

Common mistakes

Delegating a department instead of a task. "Handle marketing this week" gives the agent no way to know when to stop, so it stops at the first plausible-looking artifact.

Letting the agent choose its own sources. If the brief does not name the allowed material, the model's training data becomes the source of truth — and it will sound just as confident.

Treating a benchmark as a guarantee. The numbers on launch pages are vendor-run, on vendor-chosen tasks. They tell you the direction of travel, not what your specific job will return. Plan for the half that comes back incomplete.

Reviewing the prose instead of the process. A polished paragraph can hide a wrong source, a skipped step or a silently invented date. Check the citations and the action list first; read the prose second.

Skipping the dry run because the task looks easy. The easy-looking tasks are the ones where a wrong assumption travels furthest before anyone notices.

What to do with this today

Pick one task you were going to delegate this week anyway. Write the six fields on one page, run the dry slice, and keep the acceptance test in the same file as the output so the next person — or the next run — inherits the standard instead of reinventing it.

The agents released this month are genuinely better at finishing multi-step work than the ones from a year ago. That raises the value of a clear brief rather than lowering it: the more capable the agent, the more expensive an ambiguous instruction becomes.