Can LLM Agents Select and Engage with Biological Tools?

TEXT START: Computational tools for biological research (biological tools, or BTs) are advancing rapidly in both quantity and demonstrated capability.

The Dissection

The report is an early-warning capability audit. It tests whether LLM agents can choose biological tools and operate two of them, framing tool access as a potential gateway to broader bioweapons pathways. Its real function is to convert a vague systemic danger into measurable benchmarks, barriers, and future assessment targets.

The Core Fallacy

Relative to the Discontinuity Thesis, the central error is treating current technical barriers as meaningful safeguards rather than temporary lag defenses. The report examines whether agents can perform early-stage interaction today; the more consequential question is whether model improvement, better interfaces, tool integration, and competitive pressure will steadily dissolve the expertise and coordination barriers.

It also isolates tool use from the full system. Selection and engagement are only fragments of a larger chain involving biological design, procurement, experimentation, verification, and physical execution. That fragmentation can create false comfort. Once agents become the coordination layer across those fragments, the barrier is no longer expert knowledge. It is access, infrastructure, and control.

Hidden Assumptions

• Specialized human expertise will remain the binding constraint.
• Current model limitations will persist long enough to function as durable safety controls.
• Biological tools will remain siloed rather than increasingly integrated into agent workflows.
• Human oversight will remain competent, attentive, and structurally capable of containing autonomous systems.
• Case studies involving EVEscape and ESM3 will meaningfully represent the broader risk surface.
• Measuring present capability is sufficient to govern a rapidly improving capability trajectory.
• The relevant threat is a malicious nonexpert using tools, rather than institutions or states using agents to compress entire expert workflows.

These assumptions are not safeguards. They are points where the system currently leaks friction.

Social Function

Classification: partial truth, transition management, and prestige signaling.

The paper identifies a real capability shift: LLM agents may turn specialized biological software into interfaces accessible to actors with less expertise. That is the useful truth. But by organizing the danger around technical barriers and assessment methodology, it also gives institutions a manageable vocabulary for a problem that may outgrow institutional control. The report is a map of the first breach, not evidence that the wall will hold.

The Verdict

This is not a safety report. It is an early autopsy of the expertise barrier.

The findings, as supplied, do not prove autonomous bioweapons development. They do show the direction of travel: cognitive work once gated by specialized training is being converted into tool selection, interface management, and procedural coordination. Under the Discontinuity Thesis, those are precisely the functions most exposed to automation.

Its current barriers are hospice care for a dying control regime. The decisive question is not whether today’s agents can complete the entire biological pipeline. It is whether human institutions can preserve stable human-only control over the pipeline as agents become cheaper, more capable, and more deeply connected to biological infrastructure. The report supplies no reason to believe they can.