The runtime is the product

Field note: what three verticals taught us after the harness hype.

Back in February, everyone — us very much included — was saying "harness engineering." It was the right correction at the right time: stop blaming the model, fix the car. But it was also a generic claim. Every harness checklist we wrote in the spring could be handed to any team in any industry, and that's exactly the problem: anything that generic gets copied by August.

Then we spent a couple of quarters shipping agents into actual verticals — a bio lab, contract review, realtime dispatch — and noticed what clients were really paying for. Not the model (rented). Not the harness skeleton (week one). The remaining weeks were all one thing: runtime design that speaks the vertical.

What a vertical runtime actually is

The opinionated layer between a capable model and a domain that will fire you for a 2% error rate. Concretely: which tools exist — and which deliberately don't. What the intake asks, and in how many rounds. Where the deterministic gates sit. What the legal endings are, and who gets paged for each. Which outputs are verifiable enough to earn autonomy, and which keep a human's name on the button.

None of those decisions transfer wholesale between industries. All of them compound within one:

bio lab    → instrument MCP servers, chain-of-custody gates,
             "which freezer" is a first-class question
legal      → citation provenance, redlines as verifiable diffs,
             the runtime never paraphrases a clause
dispatch   → a 300ms latency budget that shapes literally
             every other decision in the stack

Dynamic workflows are the skeleton

If you read our dynamic-workflow note, this is the sequel. The intake + run split — everything up front, three legal endings, no questions about the salt — turned out to be the vertical-agnostic skeleton. The runtime is the muscle you attach to it, one domain at a time: the marketing runtime asks about brand voice; the lab runtime asks about sample custody; nobody's runtime asks "should I also do Weibo?" mid-run. And the harness discipline still applies — it's just table stakes now, not the headline.

The punchline

Models are rented. Harnesses are copied. The runtime — the thousand opinionated defaults about how your domain actually works — is the part that compounds, and the part nobody can fork from a blog post. The runtime is the product.

If you're building agents for a vertical we haven't seen yet, that's the most interesting email you can send us.