Structured deep-research for OpenClaw agents. Multi-pass investigation with source quality scoring, confidence levels, and open questions. For when one web search isn't enough.
A single web_search call returns headlines. For complex, contested, or nuanced topics, you need more: multiple passes, source quality evaluation, contradiction handling, and an honest confidence score.
autoresearch gives your agent a structured investigative method:
- 4-pass research loop: broad sweep → source evaluation → deep dive → synthesis
- Source quality tiers (1/2/3) to distinguish primary evidence from speculation
- Confidence score (0–10) so you know how much to trust the output
- Open questions flagged — no false certainty
Use autoresearch when:
- The topic is complex, nuanced, or contested
- You need to evaluate source credibility
- Conflicting information exists across sources
- You need synthesis, not just retrieval
- You want a confidence score and open questions flagged
Use web_search when:
- You need quick facts or current headlines
- The answer is uncontested
- A single search pass will suffice
3–5 diverse queries to map the landscape. Identifies key terms, major players, source types. Output: 10–15 candidate sources.
For each candidate source: assign tier (1/2/3), check date, note bias, flag primary vs secondary. Eliminates Tier 3 unless no better option exists.
2–3 focused queries using findings from Pass 1. Fills gaps, resolves contradictions, locates primary sources. Uses web_fetch for detailed page content.
Writes the research brief. Scores confidence 0–10. Lists open questions. Notes conflicting evidence.
| Tier | What it includes |
|---|---|
| Tier 1 | Primary sources, official government/agency releases, peer-reviewed research, official statistics, direct documents |
| Tier 2 | Major news outlets (Reuters, AP, BBC), established industry publications, reputable think tanks |
| Tier 3 | Blogs, tabloids, opinion pieces, unnamed sources, social media claims without corroboration |
| Score | Meaning |
|---|---|
| 9–10 | Multiple Tier 1 sources in strong agreement; minimal gaps |
| 7–8 | Solid Tier 1/2 consensus; minor unresolved questions |
| 5–6 | Mixed sources; some contradictions; moderate gaps |
| 3–4 | Limited Tier 1; heavy Tier 2/3 reliance; significant gaps |
| 1–2 | Scarce credible sources; major contradictions; high uncertainty |
# Research Brief: [Topic]
## Key Findings
- [Finding 1] — Source: [Tier 1/2/3 source]
- [Finding 2] — Source: ...
## Source Quality Summary
- X Tier 1 sources used
- Y Tier 2 sources used
- Z Tier 3 sources (noted where used)
## Conflicting Evidence
[Where sources disagree and why]
## Confidence Score: X/10
[Rationale]
## Open Questions
- [What we don't know yet]
- [What would increase confidence]use autoresearch to investigate [topic]
deep research on [subject] — I need source quality scores
run autoresearch on UAP disclosure news
research the history of [topic] with confidence scoring
A research brief on UAP (Unidentified Aerial Phenomena) disclosure might include:
- Tier 1: Congressional hearing transcripts, DoD/AARO reports, declassified documents
- Tier 2: Reuters/AP reporting on Congressional testimony, established defence publications
- Tier 3: Podcasts, blogs, social media claims (noted but low weight)
- Confidence: 6/10 — solid on documented testimony, low on physical evidence claims
- Open questions: nature of retrieved materials, chain of custody for physical evidence
autoresearch/
├── SKILL.md ← OpenClaw entry point
├── README.md ← This file
├── benchmark/
│ ├── tasks.json ← Sample research tasks
│ └── scorer.md ← Quality scoring rubric
├── examples/
│ └── uap_example.md ← Worked UAP research brief
├── templates/
│ └── research_brief.md ← Research brief template
└── runner/
└── run_research.md ← Research loop instructions
☕ Ko-fi · 💛 GitHub Sponsors
MIT © 2026 — free to use, modify, and redistribute.