What is VectorStep

Why VectorStep

VectorStep is built for deliberate, repeatable AI workflows, not open-ended chat. If a task has defined steps, inputs, and an outcome you want run the same way every time — an alert gets triaged, a report gets generated, a change gets reviewed — that’s the shape this is for. A capable agent is easy to demo; pointing one at that kind of work, unattended, is a different problem, and not one an agent’s own self-reported confidence solves — an LLM can be completely wrong and still say “95% confident” in a calm, well-formed sentence.

Even for personal or casual use, that’s worth having: your agents and their config live in one place, in git, and you get an actual record of what each one did and when — not results scattered across chat logs and half-remembered prompt versions. If you just want a proper way to run and keep track of a handful of agents, VectorStep already does that.

For production, a team, or anything that has to survive a security review, it’s the same tool plus what makes it defensible at that scale: the trust vector and readiness gating below, role-based auth and an audit log, calibration against real historical accuracy rather than vibes, and a security posture built assuming it’s pointed at real infrastructure. You don’t switch tools as the stakes go up — the same install scales with you.

The core idea

Every gating decision in VectorStep is built from several independent signals — a step’s own verifier, grounding checks, deterministic checks on structured output, calibration against real historical accuracy — combined into what we call a trust vector, rather than trusting one confidence number an LLM made up. Nearly every one of these checks can only ever lower a step’s confidence, never raise it — the one deliberate exception is calibration, which substitutes a measured, historical accuracy figure for the raw self-report and can move the number either direction, since it’s correcting a guess with a fact rather than applying a skeptical check. A separate readiness layer answers a different question again: not “was this run trustworthy,” but “is this pipeline, as configured today, allowed to run unattended in production at all.” Low trust doesn’t mean “fail” — it means “escalate to a human,” which is a first-class outcome, not an error path.

Pipelines are plain YAML, git-controlled — reviewable, diffable, no visual editor standing between what’s checked into source control and what actually runs. If that sounds like your kind of tool already, Quick start gets a real pipeline running against a real webhook in about ten minutes.

Read more

• Why I built VectorStep — the origin story: why “great agent demo” and “safe to run against production” turned out to be two different design problems.
• How I think about agent design — the philosophy behind keeping agents narrow and sparingly-tooled rather than broad and capable.
• Confidence & the trust vector — the full explainer for how the trust vector actually works, once you’ve got something running and want the detail.