When the question moves, revise the plan instead of stretching it

Research almost never ends with the question it started with. Halfway through, someone realizes they also need the subsidiaries, or the agent turns up evidence that the real issue is a supplier rather than the competitor everyone was worried about. That is research working as intended. The trouble starts when the plan grows to match the new question without anyone deciding that it should.

How scope drift gets in

Drift rarely arrives as one big change. It accumulates through small, reasonable-looking steps:

• The follow-up in passing. Someone adds “and check their pricing too” in a chat, and the agent folds it into the job already running.
• The promising lead. The agent finds something interesting outside the agreed scope and follows it, because following leads is what it was built to do.
• Source creep. A source that was not on the list turns out to be convenient, so it gets used just this once.
• The rewritten question. The agent quietly reframes a hard question into an easier one it can answer, and reports on that instead.

Each step is defensible on its own. Together they produce a report that answers a question nobody approved, from sources nobody vetted, at a cost nobody agreed to. The reviewer then judges it against the original question and either rejects good work or accepts work they cannot really evaluate.

Why an approved plan should stay fixed

An approval is a decision about something specific: this question, these steps, these sources. If execution can widen any of them, the approval stops meaning anything. It becomes a starting signal rather than a decision.

The fixed plan also protects the things the approver was weighing. The source list reflects judgment about reliability, licensing and confidentiality. The bounds on time and tool calls were set for the original scope. The acceptance criteria describe what a good answer to the original question looks like. A plan that stretches silently drags all of those out of date at once.

GoCux treats a research plan as three parts: the question, an ordered list of steps and an allowlist of sources. Creating a plan only records a draft; it runs once someone approves it, and running it cannot extend what was approved. An empty allowlist means no source at all, not every source. Deciding which sources may answer covers how to build that list.

Notice when the question has changed

Not every refinement is a new question. Clarifying an ambiguous term or narrowing a date range usually is not. These signals usually are:

• A new entity, geography or time period enters the scope.
• The answer will now feed a different decision than the one originally named.
• The work needs a source that is not on the allowlist.
• The acceptance criteria no longer describe a good answer.
• You catch yourself saying “technically that answers it, but it is not what we need now”.

A useful test: if the person who approved the plan read the new question cold, would they recognize it as the one they signed off? If not, it needs a new plan.

Revise without throwing work away

Revising does not mean starting over. A clean revision looks like this:

1. Pause at a clean point. Let the current step finish inside its scope, or stop it, and record what has been found so far, including the questions that have no source yet.
2. Write the new question next to the old one, with a sentence on why it changed.
3. Adjust steps and sources, marking what is new so the approver can see the difference rather than re-reading everything.
4. Update the acceptance criteria so the result will be judged against the question actually being asked.
5. Approve the revision as a new version. The original plan stays on record, along with what it produced.
6. Carry over what still holds. Findings from the first plan that answer part of the new question move across with their citations intact.

For example, the original question might be: Which of these five vendors publish a security contact? Midway, the team realizes what it really needs is which of them publish a security contact and a vulnerability disclosure policy, and whether that policy covers the product we use. The first findings carry over. The new plan adds one step and two kinds of source, each vendor's policy page and the product documentation, and the acceptance criteria now ask for the policy text quoted, not just a yes or no.

Make revising cheaper than stretching

People stretch plans because revising feels like paperwork. If revision is quick, drift loses its excuse. Keep plans short enough to re-read in a minute, show the approver only what changed, and make approval itself a single decision. Pair that with a visible ceiling on effort, as described in put a ceiling on agent spend, so a widening scope runs into a limit instead of quietly running up the bill. When the work lives on a work order, the original plan, the revision and the reason for it all sit on one record.

A quick check before you approve a revision

• The old and new questions are both written down, with the reason for the change.
• New steps and new sources are marked, and every source on the list was chosen on purpose.
• Acceptance criteria match the new question.
• Findings carried over still have their passages, and gaps are listed as open questions, as in missing evidence is an open question.
• Bounds on time, tool calls and spend have been looked at again rather than inherited.
• The original plan and its output stay on record, unchanged.

If the change is large enough that most of the old plan no longer applies, start a fresh piece of work instead; writing acceptance criteria first is the right place to begin. For more on designing research agents that keep to their brief, see research agents that hold up.