Case note
Requirements and task decomposition for AI agents
A practical buyer question: why do vague requirements make agent work longer, more expensive, and harder to accept?
An agent can only optimize against a done definition it can see. "Improve this" and "handle the ticket" do not say what is in scope, what good looks like, or when to stop. The agent then explores, takes extra turns, and produces a draft reviewers send back. The usage export shows a long run. The work record shows revisions. The cause sat upstream of the model, in how the work was scoped and split.
Clear requirements and smaller tasks make the work checkable. A useful brief names the inputs, the output shape, the constraints, and the acceptance test. Decomposition splits a large job into steps a reviewer can accept one at a time.
Synthetic example. A request says "clean up the billing module." The agent touches unrelated files and returns a wide diff that reviewers reject. The same intent, rewritten as "add validation for empty invoice dates in the existing billing form, with tests for blank and valid dates, and do not change tax calculation," produces a smaller diff against a stated test. This illustration is not a measured comparison and not a customer story.
What to inspect
- Work records: revision counts, reopen reasons, rejected outputs, and time from first attempt to accepted delivery.
- Usage logs for the same tickets: turn counts and wall time.
- The original request text stored in the ticket or chat, compared with the brief the agent was given.
- Whether large jobs were split into named steps or submitted as one open-ended run.
If the original request was never stored, you cannot audit decomposition later. Keep the brief next to the work item.
This is about scoping and splitting the work before a call is written. How to word and structure a single call is covered in AI prompting best practices: structure one prompt well. For how to read a long run once it exists, see Agent loops vs long-running work worth paying for. Why receipts beat another activity view is covered in Why AI ROI beats dashboards.
One next action
Pick one request type your team sends often. Rewrite it as a short brief with inputs, a done definition, constraints, and how a reviewer will accept the result. Use that brief for the next batch of the same request, and compare revision count and turn count against the prior batch using the same work and usage records.
If you want those records inspectable in one conversation, Overdrive helps you maximize your AI spend from inside your AI assistant. It is a remote MCP server with OAuth sign-in, and it works today in Claude, ChatGPT, and Grok Bot (setup steps). Start with a labeled sample pack and expand the lineage on each figure. Using Overdrive with your own company data is still in security review. The plan is for you to connect the tools you already use instead of uploading files, and those connectors are not available yet.
Claims and caveats
- This post claims that vague requirements drive extra turns and extra revisions, and that a written done definition plus smaller steps makes agent work easier to check against records.
- It does not claim that a better brief guarantees acceptance. It reports no revision-rate improvements, spend totals, or time reductions, and the illustration is not a measured comparison.
- The short illustration is a synthetic example. It is not customer data and it is not a result.