AI Science Workbench에는 단순한 채팅 기록이 아닌 재현성 그래프가 필요합니다.

Anthropic announced Claude Science on June 30, 2026. It describes a research workbench that can use scientific tools, create artifacts, run specialist agents, and preserve an auditable history.

Auditability is stronger when represented as a graph of immutable inputs and derived artifacts rather than a chronological chat transcript.

The graph

question
  -> dataset@sha256
  -> environment@digest
  -> analysis-script@commit
  -> result-table@sha256
  -> figure@sha256
  -> manuscript-section@sha256
                 -> reviewer-finding

Enter fullscreen mode Exit fullscreen mode

Each edge should name the operation and actor:

{
  "from": "analysis.py@4f2c1d",
  "to": "results.csv@sha256:...",
  "operation": "python analysis.py --seed 17",
  "environment": "oci:sha256:...",
  "actor": "agent:analysis-2",
  "started_at": "2026-07-14T08:12:00Z",
  "exit_code": 0
}

Enter fullscreen mode Exit fullscreen mode

Assumptions:

• raw data is content-addressed or versioned;
• tool execution happens in a captured environment;
• generated prose is not evidence for its own claims;
• reviewer agents can flag work but cannot erase provenance.

Failure analysis

Failure Detection Recovery dataset changes silently digest mismatch rerun descendants package drift environment mismatch restore image or declare variance figure edited manually missing producing edge attach source or mark manual citation does not support claim reviewer finding revise claim and retain finding parallel agents overwrite output immutable IDs create competing branches

This graph is also how I think about coding agents. I use MonkeyCode and recommend evaluating its task workflow with commits, commands, and outputs captured as durable evidence. The hosted SaaS reduces setup; the open-source self-hosted path offers more deployment control. That does not make it a scientific workbench, and I am not claiming a Claude Science integration. The useful shared principle is that long-running work should leave verifiable artifacts.

Disclosure: I’m a MonkeyCode user sharing my own experience, not affiliated with the project.

The verification path is straightforward: delete chat history from a disposable run and ask whether another person can reproduce the figure from the graph. If not, the system preserved conversation, not provenance.