On October 27, Anthropic published “Early progress on accelerating drug development with Claude”. It is not a coding story. It is the best public case study I have seen this year of what an autonomous agent does once you give it a goal, a data connection, and a way to check its work. Those are the three ingredients of Spec-Driven Development.
Read the claims with discipline first, then take the pattern.
What was actually reported#
Anthropic’s life sciences team spent several months in deep partnerships with six organizations: BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines and The Column Group. The headline results, all from Anthropic’s post and the partners’ own reporting:
- Generate Biomedicines: six scientists used Claude Code to propose redesigns of a CAR T-cell therapy component. One design was about 70% shorter than the standard version, performed better in lab tests, and produced roughly 10x more receptors.
- Divergence Labs: Claude turned research goals into a design plan and, largely autonomously, generated about 35,000 antibody designs in a month. Early tests found three binding designs and 20 that could dislodge IgE, three of which did so more than the standard medication.
- BigHat Biosciences: Claude, connected to BigHat’s lab data system, proposed changes to two antibody fragments. About 1,010 variants were tested over roughly 10 weeks. The best bound more than 10x more tightly, and one stayed stable in solution 18°C higher.
- Enveda: Claude reviewed 1,061 pages of manufacturing documentation. It caught all 11 issues human reviewers found plus seven they had missed. Human review time fell 89%, from over 25 hours to under three. A regulatory document draft took four hours versus about three months by hand.
- The Column Group: building Pleiades, a company meant to run many drug programs with a small team on Claude. Its first Claude skill, for assessing drug targets, was evaluated on 45 target-disease pairs.
- Bristol Myers Squibb: using Claude to assess early-stage clinical trial plans and estimate likelihood of success, with expansion planned for 2027.
Anthropic also acquired Mithril Medicine, a one-person company working on two rare forms of muscular dystrophy. Its founder used Claude to coordinate contracts, logistics and experiments across 11 external labs. A candidate went from idea to promising mouse results in under 18 months; human trials could start as early as late 2028 if results hold.
The caveats are not fine print#
Anthropic is candid about them, and so should we be:
- The work is early and mostly preclinical. Nothing has been tested in humans.
- Most numbers come from the partners and Anthropic itself. The post links separate technical reports, but I found no independent evaluation.
- Claude’s role in clinical trials is a plan, not a demonstrated result. BMS’s work is explicitly early-stage.
- The Enveda test was not perfect: Claude did not catch everything, and one minor error appeared in a regulatory draft.
- At BigHat, the initial metrics were misleading. More on why that matters below.
A 70% shorter CAR T component that tests better in a dish is a lead, not a therapy. Treat “35,000 designs” as throughput, not hit rate: the meaningful figure is the three binders and 20 IgE-displacing designs that came out of it.
The pattern: this is SDD with a wet lab at the end#
Strip the biology out and look at the shape of the work.
1. The goal is the spec. Divergence did not hand Claude a list of antibodies to draw. It handed over research goals, and Claude produced the design plan. That is the Spec-Driven Development loop: humans own intent and acceptance criteria, the agent owns the plan and the execution.
2. The agent is wired to the real system. BigHat connected Claude to its lab data system. An agent that can only talk is a chatbot. An agent that can read your assay results and propose the next 1,010 variants is a colleague. This is exactly why MCP and tool connectivity matter more than another point on a benchmark.
3. Skills encode the expertise. The Column Group’s first deliverable is a Claude skill for assessing drug targets, validated on 45 target-disease pairs. Same primitive you use in Claude Code to package a repeatable workflow, now holding domain judgment instead of a deploy checklist. If you have not built shared skills for your team yet, this is the argument for it: expertise written down once, evaluated against known cases, reused forever.
4. Verification is the product. Enveda’s result is only meaningful because there was a ground truth: 11 known issues from human reviewers. That is a test suite. Claude’s score was “all 11, plus seven more,” and the humans still checked the draft, which is how the minor error got caught.
5. Metrics are part of the spec. The BigHat detail is the most instructive line in the post. Standard metrics seemed misleading, and Claude helped reanalyze the data and propose new ones. Every engineer who has watched an agent optimize a green CI badge while the product got worse knows this failure. The mitigation is the same in both worlds: treat the acceptance criteria as something to be reviewed and revised, not carved in stone.
What it means for software teams#
Three takeaways, in order of usefulness.
Build your ground truth before you build your agent. Enveda could claim 89% because it had a human-reviewed baseline. If you cannot say what “correct” looks like for a task, you cannot delegate it, and you cannot measure what you delegated. Tests, golden files and reviewed examples are the real asset.
Throughput changes the bottleneck. Generating 35,000 designs is cheap; testing them is not. When an agent makes the generation step nearly free, your constraint moves to evaluation and selection. Software teams are discovering the same thing with agent-written pull requests: the scarce resource is review capacity, not typing speed.
Autonomy and oversight are not opposites. The contrast with the IDE-copilot model is stark. Nobody was tab-completing antibody sequences. The humans set direction and checked results at the edges; the agent ran the middle. That is the shape of the autonomous workflow Claude Code was built for, and a scientist (not a programmer) using Claude Code to redesign a therapeutic component is a decent argument that the terminal-native, agentic approach generalizes beyond software.
Verdict#
Real progress, honestly framed, and not yet proof of anything clinical. The takeaway for engineers is not “Claude will cure disease.” It is that the loop of goal, tool access, packaged skills and ground-truth verification works well enough in one of the hardest domains there is to be worth copying in yours. Write the spec, wire the data, encode the expertise in skills, and keep a baseline humans trust.
Sources#
- Anthropic, Early progress on accelerating drug development with Claude (Oct 27, 2026)
- Anthropic, News index
- Secondary coverage cross-checking the partner list: Digital Health News
