A Claude Agent Stack for Scientific Work

Layer one

Context must be curated before it is large

Scientific agents fail quietly when they receive a large but incoherent context bundle. The context layer should include canonical papers, protocols, dataset dictionaries, instrument constraints, lab policies, and known negative results. It should exclude stale summaries, unverified scraped claims, and private data that the run does not need.

Claude context engineering is a sibling topic in the network, so this site keeps the point narrow: autonomous science needs context with provenance. Every important source should have a title, owner, date, retrieval method, and reason it was included.

Layer two

The planner writes a protocol, not just a prompt

The planning layer turns a research question into an experimental protocol. It should define the objective, variables, controls, expected outputs, failure modes, budget, and stop conditions. For computational work, that means code paths, test data, environment files, and result schemas. For lab work, it means equipment, materials, safety constraints, and sign-off points.

A strong planner also writes the critique it expects to face. What would invalidate the result? What would count as a trivial outcome? Which measurements could be confounded?

Layer three

Tools need least privilege and typed outputs

Claude tool use allows a model to request external actions and continue from their results. In science, tool definitions should be narrow, typed, logged, and reversible where possible. A search tool, a read-only database connector, and a sandboxed Python executor belong in a different risk class than an instrument-control tool.

The practical standard is least privilege. Give the agent only the tool it needs for the next approved step, return structured outputs, and record both request and result. Human-readable logs are not enough; later evaluators should be able to replay the run.

"Claude requests to use tools"

Tool use overview, Anthropic Docs

Layer four

The artifact store is part of the science

The artifact store should preserve prompts, model versions, tool calls, code, raw results, transformed results, plots, errors, reviewer comments, and approvals. It should make negative results visible. If an agent silently retries until it finds a pleasing graph, the final report is not an honest account of the experiment.

Code execution with MCP and similar patterns are promising because they can keep code, output, and context close together. The key is to treat execution logs as evidence, not as debugging leftovers.

"reproducible code execution"

Code execution with MCP, Anthropic Engineering

Layer five

Governance decides what the agent is allowed to do next

The approval layer should be explicit. A low-risk run may allow automatic notebook execution inside a sandbox. A medium-risk run may require approval before spending compute, querying restricted data, or changing protocol. A high-risk run may never permit autonomous execution and should keep the model in advisory mode.

This is where Claude Science-style workbenches and true autonomous agents separate. A workbench improves human-led research. An autonomous stack delegates decisions. Delegation needs policy.