Guide

Parallel agent campaigns: one prompt, many agents, human review

By Atlas AI Resources · Published September 30, 2026 · How we make these guides

When the job is a huge search space — not a single ticket — one agent is rarely enough. A campaign pattern looks like this: humans write a short research brief, many agents survey in parallel, each promising lead becomes a short report, and humans decide what is worth the irreversible next step.

1. What “campaign” means here

In September 2026 Anthropic described an early life-sciences program in which Claude agents searched public DNA and protein data for unusual reverse transcriptases. Humans set the initial prompt and did the lab work. Agents gathered candidates, investigated families, and used their own judgment to flag interesting systems. One campaign spent roughly 21 hours with about 950 agents and about 210 million tokens. The agents recovered about 200,000 reverse-transcriptase clusters, promoted 17 candidate partner families, and filed 19 human-readable reports for review.

One agent noticed a repeating stretch of DNA beside an odd reverse transcriptase, treated it as worth pursuing, and filed a report. Follow-up work named the family array-associated reverse transcriptases (ART). The underlying enzyme had appeared in earlier genome papers; the write-up says Claude appears to have been the first to notice the associated repeat array and partner gene together. The biological function is still unknown — the point for builders is the workflow, not a finished biotech tool.

2. The funnel, not the swarm

Parallel agents only help if you have a narrowing path. The ART campaign is a clear funnel:

• Brief — one research question and scope (here: find interesting reverse transcriptase systems in a huge sequence collection).
• Survey — many sessions gather and classify candidates in parallel.
• Investigate — deeper look at unusual families and genomic neighbors.
• Report — short, human-readable write-ups with evidence, not a raw dump of every hit.
• Human gate — scientists pick what to test; the lab work stays human.

That matches how useful small fleets work at a smaller scale: many narrow jobs are fine; what matters is a named outcome at each stage and a person who approves anything that crosses into send, merge, purchase, or publish. A reviewer bot can recommend, but a human or a fixed rule in code makes the call.

3. Harness vs. chat windows

Anthropic notes they work in Claude Science and Claude Code — tools any scientist can use — and sometimes with a harness that coordinates many Claude sessions in parallel. The harness is software that starts work, tracks sessions, and keeps the campaign moving without a human opening every window by hand.

If you run a small Grok Bot or Cursor fleet, you already have a lighter version of the same idea: separate bots or agents for separate stages, group handoffs for visibility, and routines only after a skill is proven. You do not need a research lab’s orchestrator to borrow the shape — brief, parallel survey, short reports, human gate.

4. What humans still own

The write-up is explicit: all lab work is performed by human scientists, at ordinary biosafety levels, with no human-infecting pathogens. Physical experiments, safety, and irreversible actions stay on the human side. Agents propose and analyze; humans test.

For everyday fleets the analogue is clear. Let agents draft digests, open pull requests, or gather candidates. Keep approval on sends, deletes, payments, and anything that cannot be undone with a revert. Approval does not reverse work already done — so put the gate before the action, not after.

5. Treat hypotheses as something you study

Because agents produce so many candidates, Anthropic’s team studies the proposals themselves: what distinguishes reports they keep from ones they discard. Lessons go back into the instructions for the next campaign — teaching the system to mimic the team’s scientific taste.

That habit transfers. After a week of bot output, skim what you accepted vs. ignored. Tighten anti-jobs, report format, and “when to stay quiet” rules. A fleet improves when the brief improves, not only when you add more agents.

6. What to steal for a small fleet

• Write one brief with a named search goal and a clear stop condition.
• Split survey / deepen / report into separate jobs (or separate bots) so context stays clean.
• Demand short evidence-backed reports, not a pile of raw tool logs.
• Put a human (or a dedicated reviewer bot that only drafts for you) before irreversible steps.
• Feed accept/reject patterns back into the standing instructions.

Useful public resources

Related: Small AI fleets: handoffs · Skills before routines · Keeping AI agents safe · Cursor Cloud Agents guide