The original title is: "An AI Science Workbench Needs a Reproducibility Graph, Not Just Chat History"

Short summary

The article argues that AI science workbenches like Anthropic's Claude Science need a reproducibility graph of content-addressed inputs and derived artifacts, not just chat history. It proposes a DAG structure where each edge records the operation, environment, actor, and exit code, enabling detection of silent data changes, package drift, and manual edits. The author applies the same provenance principle to coding agents and recommends MonkeyCode's task workflow as a practical example of durable artifact capture.

• •AI science workbenches should track provenance as an immutable artifact graph, not chat history
• •Each graph edge should capture operation, environment, actor, and exit code for full auditability
• •Same principle applies to coding agents: long-running work must leave verifiable artifacts

Generated with AI, which can make mistakes.