Keywrd Scribe

Autonomy is easy. Autonomy you can leave alone is not.

Wiring a model to a publishing endpoint takes an afternoon. Leaving it running is a different proposition entirely, because every failure mode is silent: the agent publishes something it already published last week, drifts off the subject it was supposed to cover, writes confidently about a fact it invented, or repeats one narrow topic forever because that is where its research kept landing.

None of those throw an exception. An unattended pipeline with no gates does not break — it degrades, in public, at whatever rate it is scheduled to run. The whole design problem is building the checks that a human editor would have performed, as code that can refuse.

Five stages, each able to stop the run

The pipeline is explicit rather than conversational: research, cluster, plan, draft, gate, publish. Each stage produces a structured artefact that the next stage consumes, so the agent's state is inspectable at every point and a failed run can be resumed from the stage that failed rather than restarted from nothing.

Research gathers real signals about what is worth covering and clusters them by meaning, so near-identical topics collapse into one rather than becoming three thin pieces. Planning turns a cluster into a brief with a stated angle and scope. Drafting works against that brief — never against an open prompt — which is what keeps a long unattended run from drifting.

Before anything publishes, gates run: is this substantially the same as something already published, does it actually cover the brief it was given, does it clear the length and structure floor. A failed gate stops the run and leaves the artefact for inspection. It does not publish with a warning attached, because there is nobody scheduled to read the warning.

Four distinct jobs, four distinct calls

• Semantic clustering — research results are grouped by meaning rather than by string overlap, so topics that differ only in phrasing are recognised as one topic and covered once.
• Structured brief generation — a cluster becomes a typed brief: angle, scope, sections, intent. Structure is what makes the next stage checkable; an unstructured brief cannot be verified against.
• Constrained drafting — generation runs against the brief with the frame fixed, rather than against an open instruction. Most drift in autonomous writing pipelines comes from letting the drafting call decide its own scope.
• Self-critique as a gate — a separate evaluation pass scores the draft against its own brief and can reject it. Critically, this is a distinct call with a distinct prompt: a model asked to check its own work in the same breath as producing it will approve nearly anything.
• Deduplication against history — every published piece is retained as an embedding, and a new draft that sits too close to an existing one is refused before it reaches the publishing step.

413K impressions in 90 days, and nobody watching it

Keywrd Scribe operates unattended on a schedule, taking each run from research through to a published piece or a stopped run with a stated reason. On one of our own properties it produced 413,000 search impressions and 4,130 clicks in its most recent 90-day window at an average position of 7.4 — 68% of the prior full year’s traffic in a quarter of the time, with no writers involved at any point and no editor reviewing output before it went live. Every run leaves a trail: what it researched, which cluster it selected, the brief it wrote, the draft it produced and which gate passed or failed.

That trail is the actual operational deliverable. When output quality moves, the question is never “what is the AI doing” — it is which stage changed, and the artefacts answer it directly.

An agent that is safe to ignore

The outcome we judge this system by is that it can be left alone. Not that it produces something publishable — any model does that on a first run — but that on the hundredth run it is still on topic, still not repeating itself, and still stopping rather than publishing when a gate fails. A quarter that delivered 68% of the previous year’s traffic is what that looks like from the outside; the gates are what it looks like from the inside.

This is the pattern most businesses actually want when they ask for an AI agent, and the part they consistently underestimate. The model is a component. The gates, the structured hand-offs and the ability to refuse are the system — and they are what separates an agent you deploy from a demo you supervise.