Agree the harness matters more than the model, and that it's harder for data than for code. Code has tests. Data has state, environments and quality issues that shift daily. Most teams want the agent before they've got the metadata, which is backwards.
Why not build the context and capability for your team’s existing harness / agentic interface (e.g. Claude or ChatGPT/Codex). Focus on dedicated analyst/reporting tools, resources/skills, apps delivered via MCP. Not saying that effort is easy, but just that that’s where the locus of energy should be. A capability layer for your org’s (or your customer’s) existing agentic workspace.
Thanks for the comment Jason, I totally agree. It's difficult to go from individual harness to team/company harness but it is necessary and I think this is where most companies and data teams are today.
Agree the harness matters more than the model, and that it's harder for data than for code. Code has tests. Data has state, environments and quality issues that shift daily. Most teams want the agent before they've got the metadata, which is backwards.
Why not build the context and capability for your team’s existing harness / agentic interface (e.g. Claude or ChatGPT/Codex). Focus on dedicated analyst/reporting tools, resources/skills, apps delivered via MCP. Not saying that effort is easy, but just that that’s where the locus of energy should be. A capability layer for your org’s (or your customer’s) existing agentic workspace.
Thanks for the comment Jason, I totally agree. It's difficult to go from individual harness to team/company harness but it is necessary and I think this is where most companies and data teams are today.
I am actually turning this article into an online free webinar (https://community.valuefromdata.ai/c/public-events/building-an-ai-harness-for-your-data-team) where I will go through these points. If interested people can join at: http://valuefromdata.ai/join