5 Comments
User's avatar
Sharad Singh's avatar

The build-vs-power call changes once the work is regulated. Powering a horizontal agent through an MCP server is clean for distribution, but the action runs inside Claude or Copilot while the accountability stays with the buyer. So the audit record ends up split across two products. A year later someone still has to explain what the agent did and under whose policy, and half that answer lives in a system you do not control. Your data-versus-actions framing is the right cut, I just think the actions side is where the provenance bill comes due, and it lands hardest in financial services where the buyer owns the exam either way.

Tanay Jaipuria's avatar

Good thoughts!

Byblos Digital's avatar

good article! "power the agent" expands reach but hands the customer relationship to the horizontal agent - you become a backend, and backends get commoditized. worth asking where the durable value sits if you're powering.

Santosh Dodamani's avatar

In addition, you can also build an agent that on surface seems like powering an agent but actually is leveraging someone else distribution. It's why Claud Tag capability in slack does not cleanly fit within Power the Agent paradigm and rather is closer to building an agent but exploiting distribution to your leverage.

Tanay Jaipuria's avatar

Good thought!