OPERATIONALISING AGENTIC SCIENCE
Putting agentic tools to work in scientific practice.
We are still learning how to work with them. Here is what has carried across our experiments.
Model capabilities, persistent harnesses and reusable methods come together.
THE HARNESS YOU SAW EARLIER
V1 · OUR STARTING POINT
Orchestrate a scientific workflow.
V2 · WORKING TOGETHER
Augment the scientist directly.
THE COORDINATION LAYER
Keep parallel work coherent.
V3 · AUTONOMOUS CAMPAIGNS
Pursue defined goals at scale.
Feedback connects every layer: results reshape questions, methods and shared memory.
| Harness | Sessions, permissions and orchestration |
|---|---|
| Skills | Reusable methods and instructions |
| Computational tools | Scientific software through MCP or APIs |
| Execution environments | Local machines, sandboxes and distributed workers |
| Independent memory | Markdown, graphs and source records |
| Models + routers | Choose and switch the intelligence |
The workspace can span many environments. A scientist can steer the work remotely.
The harness can sit behind a familiar app.
Find and frame promising ideas.
Test those ideas at scale.
V2 frames campaigns. V3 returns evidence, surprises and new questions.
Manual review
Automated capture + AI review
What did the agent do?
What supports the claim?
Can we reconstruct the work?
Apollo’s Watcher is one safety example. You should control trace access, retention and reuse.
Replace a model or harness as needs change.
Keep skills, memory and evidence in portable forms.
Control data, permissions and the use of traces.
Use the interface that fits you. Keep the work and learning usable beyond it.
Who requested, ran and approved the work
Inputs, methods, outputs and their source
Records outside the acting agent’s rewrite access
Hugging Face investigation: agents successfully spoofed some tool calls. Capture integrity matters too.
Attestation can establish provenance and detect alteration. Scientific validity still needs its own checks.
Programmable IP ownership + rights · Lab fractionalisation
Next: connect scientific work, policy-controlled funding and attested results using a Molecule onchain lab.
IF THIS LANDS
At the pace of agentic work, with human oversight and an auditable record.
Reward credible reports of failure as well as success.
Each participant retains agency and oversight over how they participate.
WE WOULD LOVE TO COMPARE NOTES
What works today, what do you control, and where does your learning live?
The tools will keep changing. We can share what makes them useful in scientific practice.