Vertical Studio Agent Kit
A shared substrate that lets a five-person vertical team ship a production agent in a fortnight.
1 quarter to 2 wks
Time to first production agent
6 stacks to 1
Duplicate infrastructure
target 100%
Studios on the kit
The problem
Every vertical team rebuilds the same six things: retrieval, tool routing, evaluation harness, human review, tracing and rollout control. They do it slightly differently, so nothing is comparable and nothing is reusable, and platform teams spend their quarter on support.
This is the classic platform leverage play. The differentiated work in a vertical is the domain knowledge and the workflow; everything underneath it should be boring, shared and excellent.
Agent design
- A typed agent runtime where tools declare their side effects, so approval requirements are derived rather than remembered.
- An evaluation harness wired in from the first commit, with the golden set living beside the agent definition in the repo.
- A review console the vertical team can theme but not fork, so human-in-the-loop behaves identically everywhere.
- Progressive rollout with automatic rollback on eval regression, shared across every studio on the kit.
Guardrails
- A tool with write side effects cannot be registered without an approval policy attached.
- No agent reaches production without a golden set and a passing regression run.
- Traces are on by default and cannot be disabled by the consuming team.
Adoption is the only metric that matters
A platform nobody adopts is a rewrite waiting to happen. I would measure the kit purely on the share of new agents built on it, and treat every team that opts out as a product bug to interview, not a compliance failure to escalate.
What I would deliberately not build
No prompt marketplace, no visual agent builder, no universal ontology. Those are the three things platform teams build when they have run out of contact with the verticals they serve.