Implementation Checklist for Claude-Based Science Agents

1-3

Scope the job before selecting the model

First, write the research task in a form that can be reviewed: objective, domain, allowed data, expected outputs, and stopping condition. Second, choose the autonomy rung: advisory, planning, sandbox execution, or physical execution. Third, name the responsible human owner for the run.

This ordering prevents a common failure: starting with a powerful model and then retrofitting policy after the demo works.

• Define the question and acceptable evidence.
• Choose the autonomy rung and risk class.
• Name the responsible human owner.

4-6

Prepare context and tools as controlled inputs

Fourth, curate the context set with source titles, dates, owners, and retrieval methods. Fifth, define typed tools with least privilege. Sixth, create a dry-run mode that exercises the protocol without irreversible action.

For MCP-based systems, this means treating each connector as a capability with a reason to exist. Do not give a science agent a broad connector just because it is convenient.

• Curate source context with provenance.
• Define narrow tools with structured outputs.
• Provide dry-run and simulation paths.

"an open standard for connecting AI assistants"

Introducing the Model Context Protocol, Anthropic

7-9

Capture artifacts while the work happens

Seventh, record prompts, model identifiers, tool schemas, tool calls, outputs, code, data, errors, and approvals. Eighth, preserve negative results and failed calls. Ninth, make the run exportable for review.

Treat the run record as a first-class scientific artifact. If another expert cannot reconstruct the path, the agent has not produced an auditable result.

• Log model, prompt, tool, and environment metadata.
• Keep failed and negative runs visible.
• Export a human-reviewable research packet.

10-12

Review, release, and monitor cautiously

Tenth, require reviewer approval before the agent changes scope, spends money, touches restricted data, controls equipment, or publishes externally. Eleventh, run adversarial tests with seeded errors and impossible tasks. Twelfth, monitor runtime behavior after deployment.

The checklist should be re-run whenever a new tool, dataset, model, lab workflow, or publication channel is connected.

• Gate scope changes and higher-risk actions.
• Seed errors to test reviewer and tool behavior.
• Monitor protocol drift, cost, retries, and overrides.