Latency Budgeting

Latency is a feature; budget it before you build.

Agents fail user-acceptance tests on latency more often than on quality. A great answer at 12 seconds is worse, in product terms, than a B+ answer at 2 seconds for most interactive use cases. Latency must be engineered, not measured after.

When to use this skill

• The user is building a user-facing chat, voice, or agentic UI.
• The user reports an agent that "works but is slow."
• The user is choosing a model tier and is about to default to the biggest one.
• The user is adding a sub-agent, a reflection step, or a guardrail and asks how to keep it snappy.

Where time actually goes

In order of typical magnitude for an LLM agent call: