Case Studies

How Thousands of AI Agents Orchestrate Biology Research

OpenAI has used thousands of AI Agents to analyze complex biological data. The more consequential challenge is not running Agents in parallel, but coordinating deduplication, budgets, evidence provenance, and human review.

Font size

OpenAI recently described an approach that uses thousands of AI Agents to analyze complex biological data. The interesting part is not the number of Agents, but how their tasks are scheduled, deduplicated, and kept within budget.

Original article: Scaling AI research for complex biological data

Not more chat windows, but parallel exploration

A typical question-and-answer interaction is short: submit a prompt, call a model, and return an answer.

Biological data research has a different structure. It often requires multiple rounds of work:

  • Search papers, databases, and experiment records
  • Cross-analyze genes, pathways, molecular structures, or disease phenotypes
  • Combine evidence from different sources to form candidate hypotheses
  • Use the current findings to generate more targeted searches for the next round

Thousands of Agents can increase throughput, covering more documents, entity combinations, and analytical paths.

However, more parallelism does not mean a more effective search. Without clear boundaries, a system may retrieve the same papers repeatedly or repackage similar hypotheses as separate “new findings.” Worse, once an incorrect association has been repeated often enough, it can begin to resemble a credible signal.

A multi-Agent system therefore needs a scheduling layer that handles at least:

  • Task deduplication
  • Result deduplication
  • Source tracking
  • Stopping conditions
  • Budget allocation

A chat assistant focuses on the current answer. A research system must also manage the search as a whole and understand what remains after it ends.

Scheduling becomes the bottleneck at scale

As the number of Agents grows from one to thousands, operational problems emerge before model capability does.

The first is duplicate searching. Multiple Agents may work from the same keywords, abstracts, or citation trails. Without adequate shared memory and caching, the system repeatedly pays the inference cost.

The second is error amplification. If one Agent mistakes “correlated” for “causal,” others may build on the faulty premise. Shared context does not automatically correct errors; it can amplify false signals.

The third is task dependency. Some tasks require entity disambiguation first. Other analyses depend on results from an earlier stage. Some candidates are not worth pursuing until competing explanations have been ruled out.

The system cannot rely on queue length alone. It also has to maintain relationships between tasks:

search results → entity disambiguation → candidate hypotheses → evidence scoring → experiment proposals
    ↘                         ↘
     duplicate tasks merged       low-scoring tasks stopped

The fourth issue is cost. If every Agent calls an expensive model once, thousands of Agents will quickly accumulate substantial expense and latency. The scheduler must allocate budget by priority rather than distribute it evenly.

The following code demonstrates only deduplication, ranking, and budget control:

from dataclasses import dataclass
from typing import Iterable

@dataclass
class Task:
    query: str
    priority: float
    signature: str

def schedule(
    tasks: Iterable[Task],
    remaining_budget: int,
) -> list[Task]:
    seen = set()
    ranked = sorted(tasks, key=lambda task: task.priority, reverse=True)
    selected = []

    for task in ranked:
        if task.signature in seen:
            continue
        if len(selected) >= remaining_budget:
            break
        seen.add(task.signature)
        selected.append(task)

    return selected

It cannot conduct biological research by itself. A real system must also record task provenance, versions, execution status, and dependencies among results.

False positives cost more in biological research

When code is wrong, it usually produces a compile error, a failed test, or a production incident. A faulty biomedical hypothesis may not receive feedback so quickly.

A candidate target may look plausible, and its supporting citations may even exist, yet a key experiment may reveal that it has no therapeutic value. The error can then shape subsequent cell experiments, animal models, and clinical pathways.

AI Agent output therefore cannot be treated directly as a research finding. At least three layers of review are needed:

  • Source tracking: Which paper, dataset, or experiment batch supports each conclusion?
  • Evidence review: Do the underlying data support the conclusion, and is there contrary evidence?
  • Human review: Do researchers judge the candidate worth further investment?

This is where biological research imposes stricter requirements than ordinary information retrieval. A conventional search error may spoil one search. An error in biological research can waste weeks or months.

The system must also distinguish established facts from new hypotheses. An Agent can generate ideas worth testing, but it cannot present the act of generation as a completed discovery.

Agent count is not a measure of research output

“How many Agents ran?” is a poor metric for evaluating scientific work.

The better question is which candidate leads remain after the search is complete:

Metric Question to answer
Novelty Does the hypothesis go beyond existing research?
Reproducibility Do the findings remain stable when the dataset, model, or retrieval path changes?
Validation cost How much time and resources are required to move from a candidate to experimental confirmation?
Downstream value If the hypothesis holds, can it support the next experiment?

A system that produces 10,000 candidates has not necessarily discovered 10,000 research directions. If many results are repetitive, untraceable, or extremely expensive to validate, the actual yield may be low.

Evaluation should also retain a human baseline. Without one, the team knows how many tasks the system ran, but not whether it finds useful leads faster or more consistently than a researcher conducting a smaller manual search.

The results worth examining first include:

  • How many candidates remain after duplicates are removed
  • How many pass human evidence review
  • The average computational cost per valid candidate
  • How much time is saved between candidate generation and experiment scheduling

These figures are much closer to research value than the number of Agents.

Can the approach transfer to drug discovery?

Multi-Agent search fits several early-stage activities in drug discovery.

Target discovery can divide literature, omics data, and disease phenotypes into separate tasks. Literature mining can independently track drugs, targets, and experimental results before cross-checking them. Structural analysis can also explore binding paths, mutation sites, and existing molecular series in parallel.

It cannot replace wet-lab experiments. Models may shorten candidate-generation time, but they cannot substitute for cell culture, animal studies, or clinical validation.

At least these issues need to be resolved before adoption:

  • How to assess candidate novelty rather than merely text similarity
  • How to preserve dataset versions, retrieval scope, and inference traces
  • How to handle contradictory experimental results
  • How to restrict Agents’ authority to propose high-risk experiments
  • How to calculate the total cost of each valid lead
  • How to let researchers review, reject, and roll back tasks

In knowledge-dense environments with demanding validation requirements, scalability is not about whether the system can run concurrently. It is about whether concurrency remains controllable.

Editorial perspective

This approach matters not because thousands of Agents sound impressive, but because it turns research work into parallel, traceable, and stoppable tasks. Models handle local reasoning, while the scheduling layer determines which tasks merit further work, which results should be merged, and which errors should be blocked. In drug discovery, even a plausible candidate can consume weeks of experimentation, so throughput and useful output must be measured separately.

Multi-Agent systems are likely to expand first in literature retrieval, target screening, and experimental design drafting. The number of candidates reaching the wet-lab stage will not grow proportionally with the number of Agents. A well-designed scheduler, evidence graph, and human review process may matter more than upgrades to any single model.

This differs from earlier forms of chat assistants. Chat assistants generally answer one request at a time, while coding Agents can expose errors quickly through tests. Biomedical Agents operate with long validation cycles, slow feedback, and expensive mistakes. Discussions that once centered on parameter counts and benchmarks now need to include task hit rates, human review results, and experiment timelines.

The practical efficiency of this approach in drug discovery remains uncertain. The original article does not disclose the full task decomposition, deduplication rules, hit rates, or wet-lab validation data. Until those metrics are available, “ran thousands of Agents” cannot be equated with improved R&D efficiency.

Source

https://openai.com/index/scaling-ai-research-for-complex-biological-data/

Source: https://openai.com/index/scaling-ai-research-for-complex-biological-data/