docs: two-agent research-loop A/B — result + how to run

[drewstone]

A public-facing writeup of the two-agent research-loop A/B, so a Twitter/X post can link to it.

What the doc covers

• What it is — a two-agent research loop that grows one knowledge base: a worker finds web sources for the open knowledge-gaps; a verifying driver (a second agent) judges each fetched source before it's saved (on-topic relevance + near-duplicate), then gates on a readiness spec. Compared against a single-agent loop that just accumulates. Points at the real code: runTwoAgentResearchLoop (src/two-agent-research-loop.ts), createWebResearchWorker + createVerifyingResearchDriver (src/web-research-worker.ts), and the A/B harness (tests/loops/research-loop-equal-compute.test.ts).
• The result — real run, n=9 ML topics, equal compute, glm-5.2: the two-agent loop admits 2.33 fewer sources/topic at identical (1.00) coverage, 95% CI [1.78, 2.89] via agent-eval pairedBootstrap. Includes the per-topic delta table.
• The honest nuance — the win is mostly de-duplication (same canonical paper mirrored across arxiv/openreview/neurips/blogs), which is band-independent and fires on every topic; off-scope rejection is real but the minority. A cheap content-hash/canonical-URL dedup would capture most of the value without an LLM verifier.
• Threats to validity — verifier is also the judge (admitted-count is a proxy); deltas are conservative (single-agent stops early on readiness); n=2 clean offline controls is thin; glm-5.2-specific; high web run-to-run variance.
• How to run — the offline A/B (deterministic, no creds) and the live sweep (AGENT_KNOWLEDGE_LIVE=1 + TANGLE_API_KEY + AGENT_KNOWLEDGE_LIVE_GOALS |-separated, ~$0.20 for the full 9-topic sweep).

Verification

• pnpm run lint — clean (2 pre-existing warnings in wikilinks.ts, unrelated).
• pnpm exec vitest run tests/loops/research-loop-equal-compute.test.ts — 2 passed, 1 skipped (the live arm). The documented offline command ran green: [A/B @ B<=6 passes] two-agent: passes=2 junk=0 coverage=1.00 | single-agent: passes=2 junk=2 coverage=1.00.

Links the doc from the README's two-agent research loop section. Docs-only + one README line; no code changes.

[tangletools]

✅ Auto-approved PR — c4ae34fe

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-24T22:50:00Z}

[tangletools]

✅ Auto-approved PR — 367bbe4f

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-24T23:02:55Z}