Anthropic and Citrini Test 300 Claude Subagents for Autonomous Scorpion Antivenom Discovery
Anthropic collaborated with Citrini Research using an unreleased Claude model to run 300 subagents designing scorpion antivenom proteins. Out of 320,000 screened candidates, wet-lab assays confirmed 17 binders, though none neutralized the actual toxins.
Why it matters
Engineers designing autonomous agent swarms must decouple high-throughput generative screening from functional real-world validation to avoid conflating binding heuristics with viable end products.
TL;DR
- 01300 subagents orchestrated over 64 hours generated 320,000 candidates and narrowed them to 211 designs.
- 02Wet-lab validation at Adaptyv Bio and Twist Biosciences verified 17 binders across 4 of 5 toxins.
- 03Protein binding does not equal therapeutic neutralization, marking the boundary of pure in silico discovery.
Key facts
- Subagent Count
- 300 subagents
- System Prompt Length
- 35 pages
- Compute Run Duration
- 64 hours
- Initial Candidates Generated
- 320,000
- Compute & Token Bill (Anthropic)
- Just shy of $50,000
- Wet Lab Testing Cost
- Approximately $150,000
- Lab-Validated Binders
- 17 of 211 designs
Massive Multi-Agent Compute Pipeline
Anthropic granted Citrini Research direct access to an unreleased Claude model on-site after lifting standard biosecurity guardrails, subsidizing an estimated token and compute bill of nearly $50,000. The workflow relied on an extensive 35-page system prompt deploying 300 autonomous subagents. Over a continuous 64-hour runtime, the agent swarm generated roughly 320,000 protein candidates aimed at neutralizing scorpion venom toxins.
Claude filtered out candidates resembling existing proteins or known toxins, narrowing the set down to 180 designs. Anthropic researchers and Citrini then collaborated with Claude to design an additional 31 variants, resulting in 211 total proteins selected for wet-lab synthesis.
Wet-Lab Assay Results and Translation Bottlenecks
The 211 designs were sent to external synthesis and testing facilities, including Swiss biotech Adaptyv Bio and US-based Twist Biosciences, incurring approximately $150,000 in testing expenses. Out of 211 synthesized proteins:
- 17 proteins demonstrated measurable binding affinity.
- The binders collectively hit 4 out of 5 targeted scorpion toxins.
- Exactly 0 proteins bound to more than a single toxin.
Crucially, no tests proved that any binding protein actually stopped or neutralized toxin activity in biological systems. Academic researchers noted that generating target binders is already manageable through conventional high-throughput screening; the unresolved engineering challenge remains translating initial binding hits into stable, non-toxic medicines.
✓ When to use
- Running broad combinatorial searches and exploratory candidate generation across massive parameter spaces.
- Filtering millions of potential structures down to bounded sets using multi-agent heuristic pruning.
What to do today
- Separate high-throughput LLM screening stages from real-world execution and functional verification.
- Treat AI-generated binding or correlation metrics as raw hypotheses rather than verified solutions.