Field note · PDF

The loop got faster. Are you still in it?

Agentic loops now run quicker than a human can read them

Author
Jason Cruz Falzon
Published
Length
6 pages
Abstract

A Cognitive Delivery field note. Agentic loops now run quicker than a human can read them. Teams have noticed, and the instinct is to make the output prettier. This note is about why that is half a fix, and where the other half, keeping humans meaningfully in the loop, actually lives.

What the paper argues

An agentic loop is the cycle an agent runs to get work done: act, observe, decide, repeat. The phrase “in the loop” does double duty. There is the agent’s loop, which gets longer and faster as models improve, and there is the human in the loop: whether a point still exists inside that cycle where a human decision genuinely changes the outcome. Not a notification, but a decision that can go either way and is recorded. Feeling in the loop and being in the loop are not the same thing.

The human drops out through arithmetic, not negligence. The loop produces more change, faster, than the review capacity around it can absorb. Removing the human is speed without control. A richer output surface is on the right track, but it only counts if behaviour actually changes. The frame that holds is structural: detect in the loop, enforce at the gate.

Key points

  • The drop-out runs in three steps, each reasonable from the inside: the skim, the rubber stamp, and accepted-without-reading, where work ships because stopping it felt more expensive than trusting it.
  • Detection belongs at authoring time: record what was done, by which agent and model, and against which constraints, with no blocking. Enforcement belongs at the promotion gate: deterministic, default-deny for what matters, not a polite request inside a prompt.
  • A beautiful review dashboard that is still skimmed is a Level 4 control. It is not Level 5 until you can name a change a reviewer stopped because of it.
  • Six diagnostics test whether the decision point is real: named save, deterministic gate, engaged surface, evidence at hand, reversible by design, and failure-mode literacy. The named save is the one that cannot be faked.
  • Four starting actions: name a change a human stopped in one workflow last quarter; ask whether the promotion gate is enforced by tooling or by goodwill; put the detection evidence where the decision happens; and, before buying a richer review surface, decide how you will tell whether it changed behaviour.
  • The note treats one dimension closely, Review Substance, and the Level 4 to Level 5 jump that decides whether a control is real.
Written in the open

New material lands week by week

Research, the kernel devlog, news and the papers are all published here first. Get them by email, or follow the feeds.

Or follow the work directly: Research · Kernel devlog · News · CDI assessment