Sabitlenmiş Tweet
Klent
7 posts

Klent
@klentdev_
Stop your AI agents from shipping bad actions to prod. Millisecond policies, human‑in‑the‑loop approvals, replayable audit log. Works with any framework.
Global Katılım Mayıs 2026
6 Takip Edilen2 Takipçiler

And — live on HN now: news.ycombinator.com/item?id=481363…
Happy to answer questions in the thread, especially the obvious "couldn't I write this myself?" (yes, and Stripe / WorkOS / LaunchDarkly exist anyway).
English

@aistartuplabs That's the design split vs `deny`: `deny` returns a structured "blocked, here's why" so the agent retries with new params. `modify` rewrites silently but the result wraps a note (which policy fired, what changed) — agent can react or ignore, up to the loop.
English

@klentdev_ The sub-50ms thing is what makes this actually usable.
We hit the same wall building approval gates for AI execution, anything over a few hundred ms and the agent flow just falls apart. Early versions were painful.
One thing I keep running into though, when you modify a call and the agent already has a tool schema locked in, does the engine rewrite the params before they hit the tool or does the agent itself need to know the schema might change underneath it?
English

@aistartuplabs Engine rewrites the params before they hit the tool — schema stays locked, only the values change. Tool always sees valid input matching its declared signature, so agent code doesn't need to know anything was touched.
English


