A question people ask
Does human-in-the-loop actually work?
It depends which loop. The version most teams are building — a human checking the agent's output after the fact — has a measured flaw. The version that holds puts the human's judgement in before the run, not after it.
The loop the field is building
Agent acts; human approves, audits, or catches. It reads as safety.
After the production numbers turned (41–86.7% task failure in the MAST study), the field retreated from full autonomy — to the validator loop. A person downstream, reviewing what the machine already did.
The International AI Safety Report 2026 reports that reliance on AI tools can “encourage ‘automation bias’, the tendency to trust AI system outputs without sufficient scrutiny”. The checker drifts into rubber-stamping — and a human reviewing a machine that acts far faster than they can read was never a fair contest in the first place.
What the validator can’t see
Review catches symptoms. The cause is upstream.
When an agent re-opens a settled question or routes around a constraint, the root cause is rarely visible in the output — it’s the decision the agent never saw. A reviewer can catch the diff; they can’t restore the reasoning that was missing when the work was done. And every catch costs the speed that made the agent worth running.
The shape that works
Authorship upstream: the ratified why in front of the agent, before it acts.
Move the human’s judgement to where it operates structurally: the decisions the run starts from. Unl holds what you’ve settled — each call with its reasoning, superseded when your thinking moves — and serves the relevant one as unprompted relevant context at the step where it bears. Drift is prevented, not audited.
Tuned for Claude, Claude Code, ChatGPT & Cursor at launch, extending across the AI ecosystem. Connects anywhere MCP does.
Human-in-the-loop works when the human is the author, not the auditor. A reviewer downstream rubber-stamps at machine speed; a settled why upstream binds the run before it drifts.
Questions people ask
Does human-in-the-loop oversight actually work?
The after-the-fact shape has a measured flaw: the International AI Safety Report 2026 describes automation bias — the tendency to trust AI outputs without sufficient scrutiny — which erodes exactly the checking the loop depends on. Upstream shapes, where the human's settled decisions reach the agent before it acts, avoid depending on tired downstream vigilance.
What is automation bias?
The International AI Safety Report 2026 defines it as the tendency to trust AI system outputs without sufficient scrutiny. Applied to agent oversight: the longer a human approves a mostly-right machine, the less genuinely they check — the validator becomes a rubber stamp.
What does 'upstream' human-in-command look like in practice?
You settle decisions as you work — each with its why. When one bears on what the agent is doing, it's served into the agent's context unprompted, before the action. The agent keeps its speed on execution; the judgement it builds from is yours, ratified, current.
What this is
Think inside your AI world — you stay in command
Unlimitless (Unl to friends) holds what you've settled, reads what your tools are showing, and catches what's changed out in the world — and hands your AI whatever bears on the work, the moment it's needed, without you asking. The right thing, in front of the model, unprompted, with you in command of the call. So you keep moving toward what you set out to build, on top of everything you've already decided.
It plugs into Claude, Claude Code, ChatGPT and Cursor as an MCP connector. Quick to connect, in a couple of steps.
Unlimitless is open now to invited Alpha. Apply for the Beta waitlist to come in ahead of the full launch:
Alpha is invite-only · free at launch.