One fluent answer is a draft. Hard questions need parallel work you can audit. Multi-agent is that parallel. Deep research is the method, not a button you worship.
Why this matters
Pathway A taught Search, URLs, open the page, HOLD / WRONG / UNVERIFIED. That still stands. This lesson is the question with moving parts: two agencies, a rumor on X, a spec that changed this year.
A Fast reply picks a tidy story. Think spends longer on the same story. Neither automatically splits the work. Multi-agent on grok.com — official as of August 2026 — runs agents in parallel on the hardest questions. Each can show its work. Results merge into one cited answer. SuperGrok often unlocks it. Named “Deep Search / Deep Research” modes have come and gone. Do not hunt a logo. Run the method. If Multi-agent is missing, say so and do three sequential passes. Pretending you ran four agents you cannot see is not honest.
Teach
When Multi-agent earns the wait. Independent sub-problems, contested facts, or a merge that could hide disagreement. “Current public rule + what the official account said this week + what would change the plan” is a candidate. Rewriting email, summarizing your PDF, or drawing a diagram is not.
When Fast or Think is enough. One number, one file, a format transform, a critique loop. Do not spend Multi-agent quota to feel serious.
How to brief Multi-agent. The composer control sits near Fast / Think (names wander: Multi-agent, Heavy, a similar cluster). Turn it on for one message, not for the whole week. Give it a question that splits:
Mode: Multi-agent (or say if I only have Think).
Question: [one sentence]
Split the work into 3 named jobs, for example:
- Primary documents (official pages, PDFs, changelogs)
- What was said on X (official accounts vs chatter)
- Contradictions and dates
Each agent (or pass) must return: claims, URL, date on source, 6-word quote, leftover uncertainty.
Then merge:
- Agreed:
- Disagreed:
- Still UNVERIFIED:
Do not invent URLs. Do not average two dates into a third.
I will open sources myself.
Read the disagreement before you read the pretty merge. If three agents “agree” in identical sentences, you may have one thought copied three times. Ask: “Where did you actually diverge? If you did not, say so.”
Deep research is a loop, not a mood.
- Decompose. Three sub-questions you could assign to three people.
- Search. Live web and X. Primary page beats explainer. Official account beats quote-tweet.
- Quote. Six words from the page, not from Grok’s recap.
- Date. Date on the source, not the chat.
- Merge last. Agreed / disagreed / unverified. A hole stays a hole.
- You open tabs. Pathway A did not get optional because you are in Pathway B.
X is evidence of what was said. It is not a statute. If the work needs the changelog, get the changelog.
Fallback if Multi-agent will not run. Stay in one thread. Send three specialist messages, then a merge. Label them PASS A / PASS B / PASS C so the artifact can admit the fallback. Think is a reasonable mode for each pass. Search must still be on.
PASS A — primary documents only. Search the live web.
[sub-question]
Then stop.
Repeat for X-only and for contradictions. Then: “Merge A/B/C. Do not add a fourth story.”
This is not Custom Instructions, Settings personas, Grok Build subagents, or Memory. Do not ask Memory to keep the answer so you never search again. Facts go stale. Legal, medical, personnel, and anything that can hurt a person stay UNVERIFIED until a responsible human closes them.
Worked example
A coordinator asks: “Did the state change pole setback rules this year, and what are crews saying on X?”
Wrong desk: Fast, no split. Grok fuses a 2019 guide, a contractor’s “talking about 40,” and a blog caption into a new rule.
Better: Multi-agent or three passes. A finds the docket PDF — comments open, no adopted 40-foot rule. B finds crew chatter and one official “reviewing comments” post. C flags the blog caption as an older number. Merge: docket exists; 40 feet as adopted rule is unsupported. She does not put 40 feet on the tailboard. She also does not run Multi-agent on “make the brief nicer.”
Your turn
Pick a public question you can check in fifteen minutes of tabs: a spec, a fee, a product changelog, a statute, an official announcement. Not a diagnosis. Not a private HR matter. Not “who is lying about a person.” You must be able to open at least two URLs.
Open grok.com. New thread named after the question, preferably inside the B2 project if the question belongs there. Turn Multi-agent on if you have it. If the control is missing or your plan blocks it, say so in the first line and run the three-pass fallback. Both paths can pass. Invented agents cannot.
You will still open the sources. Grok will format your notes, not invent your judgments.
Exercise b5-l01-e1
| Field | Value |
|---|
| whereToDo | grok.com (Multi-agent if available; otherwise Think + Search) |
| time | 22 min |
Copy-ready prompts
Message 1:
I am in Grok Mastery (Russell Tech), Pathway B, lesson B5 — multi-agent and deep research.
Mode I believe I set: [Multi-agent / Think fallback]
Confirm in one line what you can actually do in this thread.
Question:
[one public, checkable sentence]
Decompose into exactly 3 named sub-jobs:
A) Primary documents
B) X / public statements
C) Contradictions and dates
Run them in parallel if you can; sequentially if you cannot.
For each sub-job return 2–4 claims with URL, date on source or "no date found", and a 6-word quote — or UNSUPPORTED.
Then a Merge:
- Agreed:
- Disagreed:
- UNVERIFIED:
- What would change the answer:
Do not invent URLs. Do not use Imagine.
Session stamp: today's date and "Research passes written by Grok in this thread."
Open every URL you will rely on. Take notes in your own words. Then message 2:
I opened sources. Do not change my judgments.
Format a Research Packet with:
- Mode actually used (Multi-agent / three-pass fallback):
- Sub-job A / B / C one-line findings:
- Merge you wrote (agreed / disagreed / UNVERIFIED):
- My verification table — use ONLY my notes:
Claim | Quote I took from the page or post | SUPPORTS / PARTIAL / DOES NOT SUPPORT / UNABLE TO OPEN
- Weakest claim:
- What I will actually use:
My notes:
[one short note per claim you checked]
Session stamp: today's date and "Research packet formatted by Grok in this thread from my notes."
Artifact to paste
- The three sub-job claim lists and first stamp (trim if huge; keep URLs).
- The Merge (agreed / disagreed / UNVERIFIED).
- The Research Packet including your verification table and second stamp.
- One sentence from you: whether agents (or passes) actually disagreed, or only the merge sounded complete.
Rubric
- Pass: Mode is honest (Multi-agent or fallback). Three sub-jobs exist. At least one X or public-statement line and one non-X web line, or a clear UNSUPPORTED where a pool failed. Table quotes are from pages you opened. At least one PARTIAL, DOES NOT SUPPORT, UNABLE TO OPEN, or UNVERIFIED — or you documented a real disagreement. Stamps present.
- Fail: Fast one-shot with decorative links. Fake Multi-agent (“imagine three experts”). You did not open tabs. All SUPPORTS and no hole. Private or harmful question. Packet you typed offline. Imagine used as research.
Common failure
The common failure is merge worship — the combined answer is smooth, so you skip dissent. The twin is Multi-agent on grocery lists. The third is treating an X pile-on as a primary document. Three posts are not a rule. Put them under B, not A.
What’s next
Research produces claims. A lot of desk work produces code you will not personally author for a living — formulas, regex, small scripts. Next: coding without being a developer, on grok.com, with a read-before-run rule. Not Grok Build.
