Yesterday Anthropic published how they built their internal "data agent": How Anthropic enables self-service data analytics with Claude
OpenAI published their "in-house data agent" post five months ago: Inside OpenAI’s in-house data agent
I read both - here's the result.
Where they agree
- The hard part is not writing SQL. It's finding the right table and understanding how to use it properly. Both repeat this.
- The model a commodity (they didn't say that directly but...) - the context around it is the product. Anthropic went from 21% to 95% accuracy just by adding a skill which has access to context - a knowledge base.
- (surprise, not surprise) More context doesn't necessary help. Anthropic's accuracy increased less than 1% after they gave access to all thousands of past queries (I can only imagine how much their token consumption went up 🙂).
How they differ
- OpenAI built a standalone agent. Codex and internal ChatGPT uses the agent via MCP or user can directly chat with it via web or Slack. While Anthropic built just a skill that has access to data context via some knowledge base (md files based).
- OpenAI has context in an "index" that is populated daily by a pipeline jobs while Anthropic commits the knowledge base as md files in the same repo as data model, updated in the same PR.
My take (after building a "data agent" for unstructured data, not SQL)
Anthropic's approach with skills is more "harness-native" so, it fits better to the existing toolset. I believe the whole developer experience, data context as well as the whole data platform will move under the harnesses (the best academic paper in this direction - Code as Agent Harness). OpenAI's approach seems more scalable and mature (5 months ahead?) - with a dedicated data agents, 600Tb of data and scheduled pipelines for updating the context. We will see hot it evolves.
Question for the data folks here
How far along are you on any of these agentic workflows? My understanding is that most teams are still hand-writing SQL and copy-past "context" piece by piece into Claude Code/Copilot/Codex and share knowledge in Slack.
Full side-by-side comparison table with more details and numbers in our blog post - link in comments 👇
