Deep research

Research mode is for the questions that don't have a one-paragraph answer. When you need a structured analysis with sections, citations, and a clear argument, you switch to Research mode and let Stoa run a longer pipeline on your behalf: it plans an outline, gathers evidence section by section, drafts, reviews its own draft, and revises.

It's the difference between asking a colleague a quick question over chat and asking them to write you a report by Friday.

When to use research mode

Use research mode when:

• The question needs multiple sections to answer well (background, evidence, analysis, recommendation).
• You want a real plan, not a single shot at a draft.
• The result needs to be citation-heavy and defensible.
• The work is worth minutes, not seconds.

Use the other modes for everything else: quick questions, follow-ups, lookups, drafting short text, talking through an idea.

A rule of thumb: if you'd ask a junior analyst to spend an afternoon on it, run it in research mode. If you'd ask them to find something in the next two minutes, use one of the everyday chat modes.

Starting a run

Select Research from the mode selector on the chat input and ask your question. Beside the mode selector is a second toggle:

• With approvals (the default) pauses the run at a few points so you can steer it before it goes further.
• Unattended runs straight through without stopping, answering every one of those checkpoints as an automatic yes.

Pick Unattended for a question you're confident in and don't want to babysit; leave approvals on when you want to see the plan before Stoa spends time researching it, or want a chance to send the draft back.

What happens during a run

Stoa works through a pipeline, and you can watch it move through each stage as a checklist above the answer, the same way a workflow run shows its steps: what's done, what's running, what it's waiting on, and how long each stage took.

Roughly, in order:

1. Classifying your query. Stoa works out what kind of research this is.
2. Confirming the approach (with approvals only). Stoa says what kind of research it thinks this is and lets you approve that or pick a different approach.
3. Surveying the knowledge base. A first pass to see what's actually available to answer this.
4. Drafting a research plan. Stoa breaks the question into sections, each with its own angle.
5. Confirming the plan (with approvals only). You see the section list before any of them get researched, and can approve it or cancel the run.
6. Researching the sections. Stoa gathers evidence from your knowledge bases for each section in turn.
7. Checking coverage gaps. A pass to catch sections that came out thin and go back for more evidence.
8. Writing the draft. The full report gets written, section by section.
9. Reviewing the draft. Stoa critiques its own draft for weak claims, missing citations, and unclear logic.
10. Revising the draft, if the review found something to fix.
11. Reviewing the draft with you (with approvals only). You see the full draft and the reviewer's own notes, and can approve it, send it back for another pass, or add specific instructions for the revision ("expand the methodology section," "add more on the 2024 reform"). A draft can only be sent back a limited number of times before it has to be approved or the run cancelled.
12. Final review (with approvals only). One last look at the finished report before it's saved.

Most runs take a few minutes. Long ones, deep questions over big knowledge bases, or ones that get sent back for revision more than once, take longer.

If nothing in your knowledge bases turned out to be relevant, the draft review says so directly: the report was written with no documents behind it, worth knowing before you rely on it.

Following along

You don't have to sit and watch. Once a run is going, you can leave the conversation and come back, the run keeps going in the background. When it's your turn to look at something, whether that's a checkpoint waiting on you or the finished report, it's sitting in the conversation when you return.

Stopping a run

A run in progress can be stopped from the chat it's running in. Stopping ends it there; what it had already produced is not delivered as a finished report.

Writing good research prompts

Research mode rewards specific prompts with a clear scope. A few patterns that work:

• Question + scope. "What does the 2023 funder guidance say about indirect cost recovery, and how do our last three proposals align?"
• Compare + criteria. "Compare our parental leave policy to two named peer organizations on duration, eligibility, and pay continuation."
• Brief + audience. "Brief our board on the impact of the new procurement rules, focused on what changes for grant-funded projects."

Avoid prompts that are too broad ("write me a report on compliance") or too narrow ("what's our PTO accrual rate"). Broad prompts get unfocused drafts. Narrow questions belong in the everyday chat modes.

Where to next

• Chats for the everyday modes you use day to day.
• Knowledge bases for what research mode searches.
• Citations for how to read the citations in your final draft.
• Workflows for the same step-by-step checklist, used for guided procedures instead of research.