Who babysits the agents?

The agent that worked in the demo runs under someone’s supervision in production. The supervision takes a recognizable shape: a person who notices the agent has stalled, who restarts it, who reads its output before anyone relies on it, and who absorbs the blame when it errs. And when agents start arriving with work delegated by other agents, the supervision problem acquires a second edge: a production engineer reporting from a live multi-agent deployment names the hard part:

The biggest pain wasn’t orchestration, it was trust when agents delegate to agents they don’t own. We ended up building reputation-based gating so a low-trust agent can’t delegate upward. — a production engineer, HN

The gating was built by hand because the platform supplied none. The same pattern repeats at the tooling layer: ask how operators monitor agents and the answers are dashboards, log scrapers, and vigilance. What the dashboards cannot record is the thing that matters:

Most tools record what happened (tool X was called, output was Y), but not why the agent deviated from the plan… Without causal structure in the log, you’re left correlating timestamps and guessing. — a platform engineer, HN

Correlating timestamps and guessing is a job description. It is the job the agent was purchased to eliminate, relocated to the person watching the agent.

Attention is spent, and no one accounts for it

The deeper problem is not that a human is present. A supervised system is a legitimate design. The problem is what constant supervision does to the human, because the supervision itself degrades in measurable ways, and the system never notices.

The evidence has started to arrive. A 2026 study of organizations that institutionalize AI agents on their org charts found that escalation, the operator’s main safety mechanism, changes character when the agent is treated as an employee: costly escalation rises by nearly half, and the mechanism is not diligence. The researchers’ own account: escalation substitutes for careful review and creates a record that the manager did not sign off alone. Their prescription deserves to be quoted directly:

escalation routines should “supplement—rather than substitute for—careful checking.” — Wiles, Hsu, Bedard, and Kropp, “Putting AI on the Org Chart” (2026)

This is the uncomfortable truth under every human-in-the-loop slide: the loop is operated by a person whose attention is a depleting resource, whose judgment degrades with volume and fatigue, and whose incentives include a motive that no dashboard will ever display, the motive to be seen to have checked rather than to have checked. A platform that cannot tell the difference is not supervised. It is decorated.

The industry’s answer is more chrome

Faced with a supervision problem, the tooling market ships supervision surfaces: another dashboard, another alert channel, another review queue bolted onto a framework that has no idea what the operator is there to decide. The surfaces multiply faster than the understanding. Each new pane of glass hires a person to watch it.

The defect is architectural. A monitoring surface attached to the outside of an agent framework sees what the framework emits. It cannot see what the operator needs, because that was never the framework’s problem: whether the operator has context to judge this decision, whether this is the fortieth identical approval of the hour, whether the person being asked to sign was at their desk when the work happened. Frameworks emit events. Supervision is about people.

What a system that manages attention looks like

A platform that treats operator attention as a managed resource behaves differently at every layer, and the differences are concrete enough to list.

Attention is pulled, not sprayed. The operator’s first screen is not a sea of status lights but the queue of decisions that actually require a decision, ordered by what matters, with a persistent beacon that draws the eye to the one surface where work is waiting rather than to whichever dashboard happens to be open.

A decision is an artifact, not a chat message. When the system needs a person, it parks the exact proposal, the context that produced it, and the state of the paused work into a durable case. The operator reviews, signs, and the execution resumes from where it was held. The case survives crashes, shift changes, and auditors, because it was built to be evidence, not because someone remembered to screenshot a terminal.

Escalation carries evidence. An operator who sends a decision upward annotates what they checked, and the system, because it records what the operator actually read, can verify the annotation rather than merely store it. Review becomes a supplement to judgment again. The record still shows who signed; now it also shows what the signer saw.

And the system watches the watcher. Decision volume, session length, context switches: the observable shape of operator load is read back as an advisory, flagged to the operator rather than silently accumulated. A platform that can exhaust its human without noticing is not a supervised system. It is a slow accident.

None of these are monitoring features. They are what it looks like when the management of human attention is a first-class duty of the platform, held to the same standard as the management of compute or memory.

The word for it

A system that adapts itself to conditions, including the condition of the humans operating it, without waiting to be reconfigured by hand, is self-configuring. The configuration being managed here is not knobs and thresholds. It is the boundary of responsibility between the agent and the person, drawn where judgment is needed and nowhere else, and redrawn as the system learns where judgment is wasted.

The industry will get here. It is currently hiring operators instead, which is the expensive way to learn that an org chart is not a control plane.

What is an operating system for agents? defines the layer that manages the agent. The operator surface, the durable decision case, and the evidence-verified record are part of the system design. The argument that a person belongs at the exceptions and not in the ordinary flow is made in PLRN-000.