Research

research — one question, many engines, one sourced answer

You were handed a research question. Your job is to get the real, cited answer — not one model's guess. No single search engine sees the whole picture: funding rounds live in press releases and SEC filings, reputation and traction live in X/Twitter replies, technical moat lives in a founder's blog or a HN thread, and adoption numbers surface in Perplexity's broad crawl. So you fan the question out across distinct engines that each see a different slice, run them in parallel, then reconcile their findings into one artifact where every claim carries a source. One engine is a rumor; three independent engines that agree is a fact.

Question / topic: $ARGUMENTS.

The core rule (load-bearing — read first)

Diversity beats depth from one source. Do NOT just ask Claude (or any one model) harder. Spawn engines that have access the others lack, prompt each for the slice it is best at, and keep them blind to each other so they don't converge on the same first-page results. Then cross-check: a number only one engine reports is a lead to verify, not a finding. Every claim in the final report cites where it came from — an unsourced number is a defect.

1. Frame the question — angles + a target shape

Turn the topic into 3–6 concrete sub-questions (the angles you actually need answered), and decide the output shape up front (a table of companies with funding/investors/customers? a narrative with a claims list? a comparison matrix?). Pick --depth if the caller didn't: quick (one round, ~3 engines), standard (default: all engines, one verify pass), deep (loop-until-dry: keep spawning finders until two rounds add nothing new, then a completeness critic pass). Ground once in any local context (AGENTS.md, a prior research artifact under the project's artifacts home) so you don't re-derive what's already known.

2. Fan out across engines — parallel, blind, each to its strength

Spawn these concurrently (one message, multiple dispatches). Give each the sub-questions and ask for structured findings with source URLs, not prose. Route by what each engine sees best:

Not every question needs all five. A pure market-map leans Codex + Perplexity; a reputation/sentiment question leans Grok; a technical-moat question leans Claude deep-read + HN. Pick the engines whose slice the question actually needs, but always use at least three so no single engine's blind spot becomes the answer's blind spot.

3. Perplexity (and Grok) Deep Research — the correct browser recipe

Deep Research is a dedicated mode, NOT the Computer / Control-browser / Orchestrator mode. Driving it wrong (submitting the question through computer-use) produces nothing usable. Do this:

1. agents browser start --profile <perplexity-or-grok-profile> --task research --url <site> (a signed-in profile from the injected browser list — never guess default).
2. Open a new tab / new thread, then click the Search control's caret to open its dropdown and select "Deep research" (the telescope / "In-depth reports, files, and apps" option — the same dropdown carries Search · Deep research · Learn step by step · Control browser; you want Deep research, not Control browser).
3. Type the broad brief (Perplexity's job is width — give it the whole question, not one angle) and submit.
4. Monitor to completion — Deep Research takes minutes and runs asynchronously. Poll the tab (agents browser screenshot / refs) on a bounded loop; do not declare done while it's still "researching."
5. Export the finished report (PDF / markdown / the cited sources) and save it into the run's artifact dir. Capture the source list — that's the cross-check ammunition for §4.

If the tool is blocked on credits or a paywall, that is a genuine owner blocker: post it (agents feed post "<ask>" --blocked) and continue with the other engines — never silently drop it and never fake a result. (This is exactly how a Perplexity run stalled unnoticed before.)

4. Reconcile — cross-check, dedup, and verify before you believe anything

Collect every engine's structured findings and merge on the entity/claim, not by trusting one source. For each claim:

• Agreed by ≥2 independent engines → high confidence, cite both.
• Only one engine → a lead, not a fact. Send a Claude deep-read to open the actual source and confirm or kill it. Guessing a number to fill a cell is a fabrication, not research.
• Conflicting (two funding totals, two customer counts) → surface the conflict with both sources; pick the more authoritative (filing > press > blog > tweet) and say why.

On --depth deep, run a completeness critic: "what angle did no engine cover, what claim is still single-sourced, what source did nobody actually open?" — its answers are the next round.

5. Deliver — one sourced artifact, durable

Synthesize into the target shape from §1, then write it where it survives the session, following the current project's durable-artifacts convention — never a hardcoded path. The fleet default (and this repo's own rule) is .agents/artifacts/<YYYY-MM-DD>/research-<slug>.md with any rendered HTML beside it — one dated layout, no kind-named subdirs. A project that keeps tracked research in its tree instead — e.g. a product repo whose convention is docs/research/<date>-<slug>/ — is where you promote a keepable result; --out <path> overrides the destination explicitly. Render an index.html via the artifacts skill when it's worth showing visually, with the raw per-engine outputs beside it. Every claim cites its source; a confidence column marks single- vs multi-sourced. Open the rendered result in the owner's DEFAULT browser — open <path> / xdg-open <path> locally, or scp it over and agents ssh <host> 'open /tmp/<file>' when remote (the browser they actually use; no fleet browser profile needed) — and share it if it's meant to leave the machine.

Anti-patterns

• Asking one model harder instead of many models each. The whole point is orthogonal access — Grok sees X, Codex sees the web, Perplexity goes broad. One engine repeated is not "multi-source."
• Perplexity/Grok Deep Research via Computer/Control-browser mode. Wrong mode → no usable output. Use the Search → Deep research dropdown, new tab, then monitor to completion.
• Trusting a single-sourced number. One engine's claim is a lead; verify it or mark it low-confidence. Never launder a guess into a cited-looking cell.
• Declaring a browser Deep Research done while it's still running, or silently dropping it when it's blocked on credits. Monitor to completion; park a real blocker on the feed.
• Leaving the answer in /tmp or scrollback. A research result worth having is worth promoting to the project's durable-artifacts home (.agents/artifacts/<date>/ by default) where the next agent finds it.
• Serial engines. Fan out in one message; a research sweep that runs engines one-at-a-time wastes the fleet.

Compose map

• Engines → the run skill (agents run codex|grok|antigravity|claude) and teams for a wider fan-out.
• Browser Deep Research → the browser skill; signed-in profiles + secrets for gated sources.
• Deep single-lead reading → a Claude agents run "<deep-read brief>" (or a repo-local research subagent where the project defines one).
• Render + publish → the artifacts skill (HTML) and share (shareable link).
• A blocker only the owner can clear (credits, a login) → agents feed post "<ask>" --blocked.