10 setups · each linked to the builder's own post
AI research agents: 10 real setups for research and news
Last updated · by Thomas, editor · How we verify
Pick by the job. To test a research hunch against data, start with the planet hunt in public NASA data, where the agent is told to try to disprove the idea, or an experiment loop that wakes every 3 hours and logs keep or discard. To keep up with the news, a Grok Bot routine can search the last 24 hours and send a short brief that asks for primary sources only. To watch a product or tool for change, use a changelog or new-tools watch that stays quiet on empty days. To turn research into something you can run, the bookmark demo and the idea-to-build team go one step further and build.
We chose these ten because each builder shared a public post, thread or article about the setup, and each recipe links to it. We haven't run them or benchmarked them, and we don't score or rank them. Most are prompts and schedules, and the planet hunt has the most detailed test results, which are the builder's own. Not every setup has a result to show: the planet-hunt signals are candidates, not confirmed planets yet, and after 20 experiments the autoresearch builders said it was too early to tell if it works.
What is an AI research agent?
An agent that searches sources, reads them and reports back. In these setups it usually runs on a timer: eight of the ten run on a schedule, from every 3 hours to weekly. The planet hunt and the data-analysis setup run when you start them.
What it reads depends on the setup. Several read the open web and X. The changelog and new-tools watches read official pages only. The planet hunt reads public NASA telescope data, and the data-analysis setup reads a dataset you supply. Knowing the source tells you which claims to check, and against what. The changelog watch and the data-analysis setup show the two ends.
How do research agents avoid making things up?
No setup here prevents it by itself. The fixes these builders describe are procedural: ask for the ways a result could be wrong, not whether it's real; have a fresh, read-only session audit the work instead of the one that wrote it; ask for the query behind each figure; and keep a 'primary sources only' rule.
Open any cited source yourself, because a link can exist and still not say what the agent claims. The planet hunt shows the audit approach, the data-analysis setup the query check, and the news brief the sources rule.
What goes wrong with these setups?
Each recipe's failure list names what to plan for, and several problems repeat across this group. The news brief's posted prompt has no skip line, so yesterday's story can come back until you add one (the news brief). Viral rumors get in unless the rule is primary sources only. Topics repeat unless you keep a list (the daily wonder). Interests that are too broad mean nothing gets filtered (the feed filter). Unattended loops burn money on broken configs (the autoresearch loop).
The planet hunt's most useful warning is about the reviewer. Letting the session that wrote the analysis review it is a common failure, and the builder's fix was fresh read-only sessions. An audit caught one claim that looked stronger than it was, and he removed it and redid that part. The two signals are still candidates, not discoveries, until new data and follow-up say otherwise (the planet hunt).
Which one fits your research
One card per setup, with the research job it does. Tap a title for the full recipe.
Planet hunt in NASA TESS data with Claude Code and Codex
A repeating dip in one star's NASA TESS data, tested against pass/fail rules written down first
⏱ Setup: About two weeks for the first analysis, he says: 74 separate analyses and 1,000+ scripts✋ Claims the reviewers call overstated, and anything you publish or submitAn experiment agent that wakes every 3 hours
Its own experiment log: one untested hypothesis per run, trained on a rented GPU
⏱ Setup: Half a day✋ Each keep-or-discard call, and whether it 'improved' by overfitting the evalExplore, clean and chart a dataset with a coding agent
Questions, cleaning and charts for one dataset
⏱ Setup: 30 min✋ The query behind each figure, and any decoded codes before you use themMorning AI news brief, primary sources only
The last 24 hours of AI news: 5 to 8 items, each with why it matters
⏱ Setup: 5 min✋ Repeats of yesterday's stories, since the posted prompt has no skip lineResearch feed filter: 15 sources in, only new ideas out
New items that match your stated interests, skipping what's already in your notes
⏱ Setup: ~12 hours (per the post)✋ The rejected items once, then whatever you decide to act onA daily deep-dive on one new topic
A bot that picks a new topic every morning, researches it at university depth and sends a 15-minute digest.
⏱ Setup: 15 min a day✋ The sources behind each digest, and keeping topics.md currentWatch a product changelog; stay quiet when nothing's new
New releases in an official changelog
⏱ Setup: 5 min✋ Page layout changes that break the check; the recipe says its last line catches thatDaily check for new connectors and models
A request posted publicly to Grok Bot: check daily for new tools, connectors, plugins or models and explain how to connect them, skipping empty days.
⏱ Setup: 5 min✋ How to connect anything new, since connecting is your callBookmarks → overnight demo builds
Matt Palmer's Grok Bot goes through his X bookmarks, builds a demo of one of them overnight on a branch with a preview deploy, and sends him a link each morning.
⏱ Setup: 1 hour✋ The morning demo and whether it runs; any deploy or post stays with youA weekly Grok Bot team that goes from trend scan to spec to build
Trend scans, complaints and competitors, then a spec and a small build
⏱ Setup: One weekend✋ Whether hype or star counts are standing in for proof, and any publishing or outreach
Details come from each recipe's source. Setup time is shown only where the recipe states it.
Test an idea against data
Setups that run experiments or analyses and leave a trail you can check: agreed tests, a keep-or-discard log, the query behind each figure.
Planet hunt in NASA TESS data with Claude Code and Codex
Pavel Rabtsevich, a product manager who does astronomy as a hobby, used Claude Code to search public NASA TESS data and found two transit-like signals around the star TIC 4206066, with Codex and fresh read-only sessions as independent reviewers. On Nov 6 the TESS team added the star to its list as TOI 7930 with both signals as planet candidates, with orbits within 0.29 and 1.6 seconds of the periods he published a month earlier.
An experiment agent that wakes every 3 hours
An agent wakes on a schedule, reads its own research log, forms a hypothesis, runs a training experiment on a rented GPU, keeps or discards the result, and goes back to sleep.
Explore, clean and chart a dataset with a coding agent
A workshop handout for data journalists: use Claude Code or Codex to ask questions of a database, clean messy codes, build visualizations and scrape data.
Daily news and briefs
Scheduled reads that return a short digest. Two of them keep a list, so topics and items don't repeat.
Morning AI news brief, primary sources only
A daily routine that searches the web and X for the last 24 hours of AI news and returns a short, sourced brief.
Research feed filter: 15 sources in, only new ideas out
A daily job pulls 15 curated sources, lets a model reject most items as irrelevant, checks the rest against existing notes, and saves short summaries for a weekly digest.
A daily deep-dive on one new topic
A bot that picks a new topic every morning, researches it at university depth and sends a 15-minute digest.
Watch a product or tool for changes
Checks that run every morning and only speak up when something has changed.
Watch a product changelog; stay quiet when nothing's new
A bot checks the official changelog every morning and alerts only when there's a new release.
Daily check for new connectors and models
A request posted publicly to Grok Bot: check daily for new tools, connectors, plugins or models and explain how to connect them, skipping empty days.
Turn research into a build
Research that ends in a demo or a spec. The bookmark demos land on branches, and in both setups publishing stays with you.
Bookmarks → overnight demo builds
Matt Palmer's Grok Bot goes through his X bookmarks, builds a demo of one of them overnight on a branch with a preview deploy, and sends him a link each morning.
A weekly Grok Bot team that goes from trend scan to spec to build
Nine narrow Grok Bots pass one weekly file along: a Scout finds trends, analysts mine complaints and map competitors, a Strategy Lead and CEO bot kill weak ideas, and a Spec Writer and Builder ship a small wedge.
How to choose
- Route by the job, not the brand. A hunch in public data: (recipe). A daily digest: (recipe). A watch that stays quiet: (recipe).
- Start with a watch that can say nothing. The changelog routine alerts only on a new release, and the new-tools routine skips days with nothing new. (recipe) (recipe)
- Set the filter before you schedule it. The feed filter's builder says it rejects roughly 80% as not relevant enough, and interests that are too broad mean nothing gets filtered. (recipe)
- Ask how the result could be wrong. The planet hunt's first step is to write down what would prove the idea wrong before you look further. (recipe)
- Keep builds on a branch. The bookmark demo lands on its own branch, and deploys and posts stay with you. (recipe)
Common questions
What is the best AI agent for research?
We don't rank them. Route by the job: testing an idea against data (the planet hunt, with Claude Code and Codex), a daily digest (the Grok Bot news brief), or a watch for changes (the changelog watch). Five of the ten are Grok Bot setups.
Can an AI research agent run on a schedule?
Yes. Eight of the ten here run on a timer: the autoresearch loop every 3 hours, the feed filter daily, the news brief at 08:00, the daily-wonder, changelog, new-tools and bookmark routines each morning or nightly, and the idea-to-build team's trend scan on Mondays. The changelog recipe says to run it once to record the current version, then schedule it.
Do research agents make up sources?
In general they can, and these recipes don't claim otherwise. Their own warnings are about confident numbers and rumors slipping past a 'primary sources only' rule. So open each cited source yourself, check that it says the claim, and ask for the query or file behind any figure.
What does it cost to run these?
The recipes state different costs. Four Grok Bot routines are listed as included in the plan. The autoresearch loop needs GPU and model usage, and the recipe's first step is a hard spending cap at the GPU provider. The data-analysis workshop spent $23 of tokens across all attendees. The planet-hunt builder mentions two Claude Max x20 plans and a ChatGPT Pro plan, not a total. We haven't priced any of them.
Which agents do these run on?
Five run on Grok Bot. The planet hunt uses Claude Code and Codex, and the bookmark demo uses Grok Bot with a Cursor agent to build. The data-analysis setup uses Codex or Claude Code. The autoresearch and feed-filter setups are listed for any agent with a scheduler.
Is there an open-source AI research agent?
Some pieces here are open. The autoresearch loop runs karpathy/autoresearch, an open-source project on GitHub. Simon Willison's data-analysis workshop handout is free. The planet hunt's code is published on Zenodo under MIT, with the data under CC BY 4.0. The idea-to-build team can use REA, an open-source MCP and CLI. You still pay for the model, GPU or agent plan that runs them.
Sources and credits
- AI Agent Rating: Best AI Research Agents in 2026: Routing by job in the quick answer, the 'where the evidence comes from' split, their 'how we're rating these' note of assigning no scores of their own, and the 'what is an AI research agent' question. We didn't take their product list.
- News Digest AI: Best AI Research Tools in 2026: Choosing by research job rather than by 'smartest', and a list of research failures to look for. We didn't take their tool list, scorecard or verdict.
Get the next money recipe. Free.
New recipes every week, plus the top 10 agent setups pack. One email, no spam.
Unsubscribe anytime. Privacy
More guides: Make money with AI agents · AI voice agent business · AI agent agency ideas · Claude Code use cases · Grok Bot use cases · Claude Code vs Codex · AI agents for small business · AI email agents · AI code review agents · Parallel coding agents · AI agents that save you money · AI refund agents · Autonomous agents that made money · AI agents for content creators