agents in biology need better roads, not just better models
reflections on why biological AI agents need better infrastructure, not just stronger reasoning.
what the Anthropic piece made clear
Anthropic's essay on paving the way for agents in biology made me think less about model size and more about roads. software agents seem to advance faster partly because software environments already have roads: package managers, clean APIs, test loops, execution layers, version control, and repeatable interfaces.
biology is different. many workflows still move through browser clicks, inconsistent metadata, scattered databases, awkward APIs, and expert-only conventions. even when the task sounds simple, the path through the information can be fragile.
models still need deterministic support
the viral sequence retrieval example stood out to me. strong models could reason about the task, but they were unreliable when the workflow depended on exact retrieval. accuracy improved dramatically when a deterministic retrieval layer was added.
that feels like a bigger lesson for scientific AI. many research workflows need near-perfect reproducibility. if an agent retrieves the wrong sequence, misses metadata, or silently takes a different path, the result is not just inconvenient. it can break trust in the whole system.
better biological infrastructure
better models alone are not enough. biology needs more agent-friendly infrastructure: reliable retrieval, inspectable steps, standardized metadata, safer execution layers, and tools that make each decision visible.
this connects directly to what i want to work on: biomedical AI, research tooling, and systems that turn data retrieval into something reproducible and inspectable. if agents are going to help with biology, we need to build the roads under them as carefully as we build the models.