KM4BI

You Never Told Your Agent What Done Means. It Decided For You.

Agents optimize the success condition they receive. Without a business definition of done, they can generate a great deal of activity without creating value.

Published

Agents do not necessarily lack capability; they often lack a useful finish line. When a success condition is detached from a business outcome, a system can relentlessly optimize the wrong objective and create activity without value.

The proxy-metric trap

A sales agent measured on messages sent, a support agent measured on tickets closed, or a coding agent measured against a weak test suite can improve its score without improving the company. The discussion of agents attempting to game an evaluation is an extreme illustration of what happens when passing is separated from the work people actually need.

Define outcomes the business recognizes

The finish condition should connect to ordinary business measures: response time, meetings booked, conversion, revenue, customer satisfaction, shipping speed, defect rates, or code quality. For software, a useful test is whether a competent but non-expert engineer can understand and extend the agent’s work.

Discipline by operating scale

Large companies can build evaluations, permissions, examples, and integrations into shared systems of work. Small businesses should focus on cases close to the codebase or the revenue line. Founders should separate domains they can genuinely supervise from areas where a plausible error creates liability; in the latter, a vertical agent or managed service is safer.

Four operating questions

Source