MetaHarness: research vertical

A ready-made multi-agent research harness for Claude Code. One command scaffolds a project with six specialized agents — scout, web-searcher, source-grader, synthesizer, fact-checker, citer — wired through tiered model routing (haiku for fan-out, sonnet for synthesis, opus for adversarial fact-checking) so you can run deep, multi-source, cited research the moment npm install finishes. It is not a hosted SaaS, a search API, or a finished UI — it is the prompts, agent definitions, hooks, and Claude Code settings you'd otherwise spend a weekend writing.

Quickstart

npx @metaharness/research@latest my-bot
cd my-bot && npm install && npx harness doctor

harness doctor checks your Node version, Claude Code install, and required MCP servers. When it reports green, open the project in Claude Code and ask: "Research the state of robotic foundation models in 2026 with at least 12 sources."

What you get

• agents/scout.md (haiku, Tier 2) — decomposes the question into 5-10 sub-queries and seeds the search frontier.
• agents/web-searcher.md (haiku, Tier 2) — fans out parallel WebSearch + WebFetch calls, one per sub-query.
• agents/source-grader.md (sonnet, Tier 3) — scores each source on authority, recency, and primary-vs-secondary, drops the weakest.
• agents/synthesizer.md (sonnet, Tier 3) — produces the draft report with inline citation anchors.
• agents/fact-checker.md (opus, Tier 3) — adversarially re-reads each claim against its cited source and flags drift.
• agents/citer.md (haiku, Tier 2) — normalizes citations into a single bibliography (APA + URL + access date).
• .claude/settings.json with WebSearch, WebFetch, Read, Write pre-allowed; Bash(rm:*) denied.
• .claude/commands/research.md — slash command /research <topic> that fires the full six-agent pipeline.

Advanced

Validate the scaffold's structure before you trust it on a real project:

$ npx harness validate
PASS  agents/scout.md          (frontmatter ok, tools allowed)
PASS  agents/web-searcher.md   (frontmatter ok, tools allowed)
PASS  agents/source-grader.md  (frontmatter ok, tools allowed)
PASS  .claude/settings.json    (schema valid)
PASS  .claude/commands/research.md
6/6 checks passed

Run the pipeline headlessly from CI or a cron job, with a per-run budget cap:

$ claude -p --plugin-dir my-bot --max-budget-usd 1.50 \
    "/research impact of EU AI Act on open-weight model releases"
[scout]        decomposed into 7 sub-queries
[web-searcher] 38 sources fetched, 6 dropped (paywall/404)
[source-grader] kept 22 of 32 sources (avg authority 0.74)
[synthesizer]  draft written: report.md (2,840 words, 22 citations)
[fact-checker] 1 claim flagged: "EU AI Act enforces from Aug 2025" -> source says Feb 2026
[citer]        bibliography.md written (22 entries, APA)
$ harness cost
total: $1.21 / $1.50 cap

Regenerate one agent without re-scaffolding the whole project:

$ npx harness regen agents/fact-checker.md
regenerated agents/fact-checker.md (sonnet -> opus, prompt v2)

FAQ

Q: Do I need an Anthropic API key, or does Claude Code's subscription cover it? A: Either works. If you launch the pipeline from inside Claude Code interactively, your subscription covers it. If you run it headlessly via claude -p, you need ANTHROPIC_API_KEY in your environment.

Q: Can I swap WebSearch for a different search provider (Tavily, Brave, SerpAPI)? A: Yes. Edit agents/web-searcher.md — replace the WebSearch tool reference with an MCP server of your choice, then update .claude/settings.json to allow the new server. The rest of the pipeline doesn't care which search backend you use.

Q: Why six agents instead of one big prompt? A: Cost and accuracy. Fan-out search and citation normalization are cheap haiku work; adversarial fact-checking is where opus earns its keep. A monolithic prompt would either burn opus tokens on URL fetching or use haiku on claim verification — both are wrong trades. The vertical encodes the right routing.

License

MIT. Built on metaharness (https://www.npmjs.com/package/metaharness).

Deep-dive

Full explainer gist: https://gist.github.com/ruvnet/7ad69e82554b1f144aa4f5b34498536b