Research agents

Research agents

Standard research phases answer a fixed question. But some investigations don’t fit a template — you want to trace a thread across the web, check several sources, and come back with a written briefing. That’s what AI research agents are for.

An agent runs a step plan: a list of steps you author, each calling a web tool. The agent executes the steps one by one, gathers what each returns, and writes a single synthesized Markdown report at the end.

The tools an agent can use

A step plan can call three tools:

• Web search — search the open web for a query and return the top results (titles, URLs, and descriptions). This is the workhorse: start every investigation with a few targeted searches.
• Fetch a page — load a specific web page and extract its readable text. Use this to read a promising result in depth rather than judging it from the search snippet.
• Take a screenshot — capture a visual snapshot of a page (for example, a competitor’s pricing page or landing page) so the report can include what the page actually looks like.

When a tool can’t be used (for example, the screenshot tool isn’t enabled in your workspace), the step fails with a clear message — the agent never invents a fake result to paper over the gap.

Writing a step plan

A good plan is short, specific, and ordered. For example, to research the competition for a newsletter analytics tool:

1. Web search — newsletter analytics tools market 2026
2. Web search — "newsletter analytics" pricing comparison
3. Fetch — the pricing page of the most interesting competitor found above
4. Screenshot — that competitor’s pricing page

A few tips:

• One question per step. A search for best email tools small business pricing reviews returns mush; three narrow searches return usable results.
• Use later steps to go deeper. The strongest pattern is search → pick a promising result → fetch or screenshot it.
• Keep plans under about ten steps. If you need more, split the investigation into two plans.

Running a plan

1. Open your idea’s Research section and go to the AI research agent panel.
2. Add your steps, choosing the tool and its inputs for each.
3. Save the plan — it’s stored with your idea so you can tweak and re-run it.
4. Click Run.

While the plan executes, you see live progress per step: which step is running, which completed, and which failed — updated as it happens, so you can watch the agent work rather than staring at a spinner.

When every step finishes, the agent synthesizes the outputs into a Markdown report with a sources section. Download it, keep it in your idea, or use it as the basis for a canvas or a pitch.

When to use agents vs. standard phases

Use standard phases when a phase already answers your question — trends, market sizing, competitor analysis, and the others are purpose-built, structured, and feed your PMF score automatically.

Use agents when:

• The question is bespoke — “who launched a similar product in Europe last year and how did they price it?”
• You want to combine several sources into one narrative briefing.
• You need visual evidence, like a competitor’s page.
• You’re exploring an unfamiliar space and want to see where the trail leads.

The two approaches complement each other: run standard phases to build your structured evidence base, then deploy an agent for the questions that don’t fit the templates.

Note

Agent reports are written from the live web results the tools returned. Treat them like any research: verify the claims that matter before betting on them, and note that agent output isn’t fed into your automated PMF signals the way standard phases are.

If the agent fails partway, the report includes the steps that failed and why — so a partial investigation still shows you what it found before it stopped.