Latency is a feature: designing agents people trust.

There’s a quiet assumption baked into most agent demos: faster is better. Watch the tokens stream, watch the tool calls fire, watch the task complete in seconds. It’s a compelling show. It’s also, for a surprising number of real workflows, exactly the wrong instinct. When an agent acts on something that matters — money, records, a customer’s account — the feeling you’re designing for isn’t speed. It’s trust.

And trust has a pace. A system that pauses, shows its work, and asks before it commits feels dependable in a way that a blur of instant activity never does. Used deliberately, latency stops being a cost to minimise and becomes a feature to design.

Speed is cheap; confidence is expensive

The hard part of an autonomous system was never making it act. It’s making people comfortable letting it. A human operator extends trust the same way they would to a new colleague: by watching how it behaves, seeing it ask when unsure, and confirming that it stops at the edges it’s supposed to respect. None of that is instantaneous. All of it is the point.

An agent that pauses at the right moment earns more trust than one that finishes first.

The checkpoints that matter

Designing a supervised agent is mostly about deciding where it slows down. A few principles guide where we place the friction:

• Visible reasoning before irreversible action. Anything that can’t be undone — a payment, a deletion, an external message — gets a moment of legible explanation and a human confirmation. The pause is the safeguard.
• Confidence-aware pacing. When the model is sure and the action is cheap, move fast. When it’s uncertain or the stakes are high, slow down and surface the doubt. The system should feel calmer exactly when it should be.
• A real stop button. Oversight isn’t a checkbox at the end. It’s a continuous ability to inspect, interrupt, and override — designed in from the start, not bolted on after an incident.
• Honest progress. Showing what the agent is doing, and why, turns waiting into reassurance instead of anxiety. Silence reads as a hang; narration reads as competence.

Designing the wait

If a step takes time, that time is part of the interface. A spinner says “something is happening, possibly forever.” A line that says “checking three sources, then drafting for your review” says “this is under control.” The same number of seconds can feel reckless or careful depending entirely on what you choose to reveal during them.

This isn't an argument for artificial slowness. It's an argument for honest pacing — fast where speed is safe, deliberate where it isn't, and transparent throughout.

Dependable beats impressive

The agents that get adopted — the ones still running a year later — are rarely the fastest in the demo. They’re the ones whose users learned, through repetition, that the system would stop when it should, ask when it wasn’t sure, and never quietly do the wrong thing at speed. That reputation is built one well-placed pause at a time.

So before you optimise an agent for latency, ask what it’s acting on. If the answer is anything a person would want to double-check, the most valuable thing you can engineer isn’t a faster response. It’s a moment of earned hesitation — and the visible confidence that comes with it.