Day 2 · Advanced Automation Track · Lesson 08

Multi-Step Research Workflows

Course Target: Put an agentic browser to work on real research — running multi-step discovery loops across many sites, keeping track of what it finds, and synthesizing the results into a verified comparison table you can actually rely on.

Lesson 07 introduced agentic browsing; this lesson turns it into a research method. A single search returns a page of links; a discovery loop works across many sources toward a goal, remembers what it has found, and assembles the results. The catch: the agent’s tidy final table can hide wrong, stale, or invented cells — so provenance and verification are the whole game. A full discovery loop needs an agentic browser, but the scoping and verification skills here work with any approved AI that can search the web — use what your institution allows. This pairs with Lesson 09 (extraction) and Lesson 10 (verifying AI output).

Start from your task

A concrete goal for this lessonYou need to find grant opportunities for a new program — scattered across a dozen funder sites — and put them in one comparison table with deadlines, amounts, and eligibility. By the end you’ll know how to scope an agentic discovery loop to do the legwork, and how to verify the table before you trust a single row.

Learning Objectives

Scope a multi-step research loop for an agentic browser, setting boundaries (sources, depth, stop condition) so it would terminate with useful results.
Explain how an agent carries state across sites, and build in the checkpoints that keep a research task from drifting, duplicating, or going stale.
Synthesize and verify a multi-source comparison table — every cell traceable to a source and checked before use.

A note on what "able to" means hereThese activities are your first supervised attempts at each skill — a working draft and a repeatable method, not full mastery.

Key Terms

Select any card to flip it. Cards are keyboard-operable — press Enter or Space.

Course Outline & Lesson Modules

If you remember one thing: A discovery loop needs a stop condition. Without "when to stop and what counts as enough," an agent will crawl forever, drift off-topic, or quietly give up — and you won’t know which.

The picture

1. Goal + stop condition   (what you need; when to stop)
2. Search a source
3. Collect findings        (keep each source link)
4. Update memory           (what's found / what's left)
        ↑ repeat 2–4 until the stop condition is met
5. Synthesize into a table
6. Verify each cell against its source

A single search is one shot; a discovery loop repeats — search a source, collect findings, note what’s left, and continue until the goal is met. To make it useful you scope it up front:

Sources — which sites/databases (and which to avoid: logins, paywalls, anything with student data).
Depth — how far per source (top results, first pages) so it doesn’t spiral.
Stop condition — e.g., "10 qualifying results" or "all listed sources checked."
What to collect — the exact fields you need, each with its source link.

When a loop isn’t the right toolSome research is better done by hand — a focused library-database search, a single authoritative source, or anything requiring a login or licensed databases. Reach for a discovery loop when the answer is spread across many open pages.

A reusable instruction template

Paste and adapt this when you launch a loop — it bakes in the scope and the guardrails:

Search [these 2–4 sources] for [what you need].
For each result, collect: [field 1], [field 2], … and a direct source link.
Stop at [N qualifying results] or when all sources are checked.
Do not log in or open paywalled pages, and take no other actions.
Flag duplicates and anything past its deadline; note each finding's date.

Caution — the agent reads untrusted pagesAn agentic browser visiting many sites is reading content you don’t control, which can carry hidden instructions (indirect prompt injection — Lesson 11). Keep it to gathering and drafting; don’t let it take consequential actions mid-loop, and don’t let it log into your accounts or paywalled sites during a research loop — that exposes credentials and can violate a site’s terms.

Quick check: you ask an agent to "research funding options" with no limits. What goes wrong?

No stop condition and no scope — it may crawl endlessly, wander into irrelevant or untrusted sites, or return a thin result with no way to tell if it actually finished. Define sources, depth, a stop condition, and the fields to collect first.

Cumulative Check

You send an agentic browser to compile a grant comparison table. Name one risk at each stage — loop, memory, synthesis — and a control.

Reveal model answer

Loop: no stop condition → it crawls forever or drifts; set sources, depth, and a stop condition. Memory: duplicates and stale entries → ask it to de-duplicate and date each finding. Synthesis: invented/overstated cells → require a source link per row and verify deadlines, amounts, and eligibility before you trust the table.

Hands-on Workspace Modules

Activity 01

Scope a Discovery Loop

Turn a vague research goal into a bounded loop: sources, depth, stop condition, and the exact fields to collect — so the agent finishes with something useful.

sources + depth set;
a clear stop condition;
fields to collect, each with a source link.

Open Activity 1

Activity 02

Track the State

For a multi-site research run, decide what the agent must remember and where it could drift, duplicate, or go stale — and add the checkpoints that catch it.

what to remember named;
drift/dup/staleness risks spotted;
a checkpoint for each.

Open Activity 2

Activity 03

Synthesize & Verify

Check an AI-compiled grant table against its sources, flag wrong/stale/unsupported cells, add the opportunity it missed, and produce a defensible table.

each row checked against its source;
bad cells flagged + missing row added;
a verified, defensible table.

Open Activity 3

Wrap-up

Reflect (2 minutes)Think of a research task you do by hand — a literature scan, a funding search, a tool comparison. What would its discovery loop’s stop condition be, and which cells would you always verify?

Model what you teach: disclose your AI useWhen AI gathered or synthesized sources for you, note it — and keep the source links so others can verify. A one-line note is enough.

RecapLoop: scope it with sources, depth, and a stop condition. Memory: de-duplicate and date findings to avoid drift and staleness. Synthesis: a source link in every row, and verify the cells that matter before you trust the table.

Sources & Further Reading

Primary Official documentation for your agentic browser / research tool — its capabilities, limits, and any rate or access constraints (verify current features before relying on them).
Primary The original source page for every finding — provenance is the authority; a claim is only as good as the page it links to.
Primary FERPA, 20 U.S.C. § 1232g; U.S. Dept. of Education, Student Privacy Policy Office (studentprivacy.ed.gov) — if any research step touches student or institutional data.

Agentic browsing features change quickly and vary by tool — confirm current capabilities and respect each site’s terms of use and your institution’s approved-tools list.