A 3-agent crew for AI/ML researchers and curious devs โ crawls arXiv/Semantic Scholar, summarizes papers in your domain, tracks citations and follow-ups.
Staying current in a fast-moving research field โ large-language-model interpretability, retrieval-augmented generation, biology + AI โ means reading 5-10 papers per week, every week. The standard tooling (Connected Papers, arXiv RSS, Semantic Scholar alerts) gives you a firehose; what you actually need is a curated trickle plus enough synthesis to know whether the field is moving and where.
This 3-agent research rabbit hole team is the curation + synthesis layer. Scout crawls arXiv and Semantic Scholar for your interest tags and ranks by citation velocity. Notes writes 200-word executive summaries per paper โ problem, method, key result, what's missing. Survey synthesizes the week's papers into a domain map: clusters of similar work, who's at the frontier, what gaps look exploitable. You get a Monday digest with the papers worth reading and a snapshot of where the field is. The crew is built for one researcher tracking 1-3 sub-domains in depth, not a literature-review machine for a whole university department.
Crawls arXiv, Semantic Scholar, and Google Scholar daily for new papers in your interest tags.
Synthesizes paper findings into a one-page domain map; tracks who's working on what.
Writes 200-word executive summaries per paper with the exact takeaway you need.
Configure your interest tags (e.g. 'sparse autoencoders', 'long-context retrieval', 'RLHF alternatives').
Scout queries arXiv + Semantic Scholar nightly, dedupes against papers you've seen, ranks by citation velocity.
Notes generates a 200-word executive summary per paper: problem, method, key result, what's missing.
Survey synthesizes the week's papers into a domain map โ clusters of similar work, who's at the frontier, gaps.
You get a Monday digest in Telegram with the 5-10 papers worth reading and Survey's domain map.
If a paper cites work you read 3 months ago, Scout flags it ("this is a follow-up to X you read on date Y").
Quarterly: Survey writes a 'state of [topic]' overview pulling threads from the quarter's reading.
Monday research digest (sparse autoencoders, 2026-04-29): - Scout: 12 new papers indexed (3 arXiv, 9 SemanticScholar). 7 cited >5 times in last 30 days, 5 cited 0-2 times. - Notes (top 3 summaries): โข Anthropic 2026 โ "Scaling SAE features to 100M" โ pushes feature count 4x prior SOTA, training cost 3x. Key result: feature interpretability degrades non-linearly past ~16M, suggests architectural change needed. โข DeepMind 2026 โ "SAE features as routing" โ uses SAE outputs for MoE expert routing. 2.1% perplexity improvement; orthogonal to interpretability work. โข MIT 2026 โ "Concept erasure via SAE features" โ practical method for unlearning specific concepts; ~6 hr per concept on a 70B model. - Survey: This week's frontier moved toward 'features as substrate for downstream tasks' (3/12 papers). Anthropic + DeepMind racing on scale; MIT carving an applied niche. - Action items: read top 3 in full, watch the 'features-as-substrate' thread next week.
Connected Papers is a graph-traversal UI โ great for exploring around a paper you already know. Elicit answers research questions across a corpus. This crew is daily-curation focused: it knows your interest tags, ranks by citation velocity, writes summaries in your voice, and tracks follow-ups across months. Different jobs; we use Connected Papers ourselves alongside this crew.
arXiv (cs.*, stat.ML, q-bio.*, configurable), Semantic Scholar (S2 API), Google Scholar (light scraping with rate-limit respect). You can add custom sources โ IEEE, OpenReview, ACL anthology โ with a small adapter. The default config is tuned for AI/ML; biology/physics/social-science work fine but may need tag tuning.
The default SOUL.md asks Notes to surface 'what's new', 'what's the empirical result', and 'what's missing' โ which catches most contributions but can miss nuance in highly mathematical papers. Notes flags 'low confidence' summaries for papers it doesn't fully understand; you read those in full. In our testing on 200 sampled AI/ML papers, the summary correctly identified the main contribution 91% of the time.
It uses arXiv preprints and S2's open-access copies where available; falls back to abstract-only for paywalled work. If you have institutional access, you can wire your library proxy and the agent will use it. Default is open-access only.
Notes is the heaviest โ Sonnet/Opus class for accurate summaries (Haiku-class will hallucinate confidently). Scout is light, runs fine on Haiku or Llama 3.3. Survey is medium, Sonnet recommended. You set provider per-agent in each SOUL.md.
Three SOUL.md files (Scout, Survey, Notes), an AGENTS.md coordination file, arXiv + S2 + Scholar adapters, a citation-graph database snippet (SQLite), and the setup README. Drop into your OpenClaw agents/ folder; runs nightly via cron or whenever you trigger a digest.
Get 3 AI agents working together โ pre-configured, two Terminal commands to deploy.
7-day money-back guarantee ยท One-time payment, yours forever
AI competitor intelligence team that tracks pricing, feature launches, keyword rankings, and review sentiment so solo founders ship informed product calls each week.
AI product launch team that coordinates the timeline, writes the content, optimizes SEO, schedules social, and tracks competitor reactions โ built for solo founders and lean PMs.
AI data analytics team that pulls Stripe/GA4/database metrics daily, flags anomalies before they become incidents, and benchmarks your numbers against industry priors โ built for SaaS founders without a data team.