Every agent demo runs at the speed of execution. Production runs at the speed your org can actually digest what happened.
I've been calling this metabolic rate: how fast can your organization absorb agent actions, including the errors, disputes, exceptions, and ambiguities those actions inevitably produce? The GSA Inspector General put it plainly: bots "can perform thousands of read, write, and deletion actions at high speed, which can make logic and processing errors hard to identify before serious damage occurs."
Risk accumulates in the gap between execution speed and digestion speed. Organizations with high metabolic rates already have the infrastructure: review artifacts, escalation paths, structured dispute processes, fast rollback. Coding scales agents well because repos are review-shaped by default. Payments scale because chargebacks and fraud detection were priced in decades ago.
So before you deploy an agent, the capability question is the wrong one. Ask whether you can catch and repair its mistakes at the speed it makes them.
High metabolic rate looks like:
- Stable case identifiers linking actions to outcomes
- Review artifacts native to the workflow (diffs, claim files, invoices)
- Known exception paths with clear escalation
- Rollback or compensation when state is mutated incorrectly
- Disagreement already has a forum: code review, dispute, audit
Low metabolic rate looks like:
- Work moves through DMs and personal spreadsheets
- Approvals happen at a different point than policy says
- Exceptions depend on one person knowing "how this really works"
- Systems record state changes but not why they were allowed
- Handoffs visible only after something breaks
Why it degrades under load: Endsley and Kiris (1995) documented that passive supervisors of automation lose the ability to intervene after failures. Too many low-stakes approvals train people to stop reading. Metabolic rate drops precisely when execution volume rises.