Cost and Latency Are Agent Constraints

Agent loops have a cost curve. Retries, tool calls, and long context add up.

Latency has a trust curve. Users abandon waits. They also abandon wrong speed.

A support org let an agent retry tool calls five times to avoid human handoff. Average handle looked automated. Customers waited. A chunk of tickets got worse answers after the retries burned context. Unit cost spiked on the hard cases.

The claim

Design max spend and max wait per task type. When either trips, hand to a person. Do not optimize for automation rate alone.

Put numbers on the path

• Max tool calls per ticket class.
• Max seconds before human offer.
• Max cost per successful resolution.
• Cases where human-first is required.

They cut retries to two on standard tickets and sent edge cases to humans sooner. Automation rate fell a little. CSAT and cost per resolve improved. Thier CFO cared about that pair more than the vanity rate.

Ask before you scale

• What does a wrong cheap loop cost in cleanup?
• Where is slower human work already better?
• Which metric will we refuse to game?