X · Governance & ObservabilityEmerging★

Agent Withdrawal Criteria

also known as 撤退基準, Pre-Agreed De-Adoption Thresholds, Agent Exit Criteria, Expiry-Date Design

Agree before production on the quantitative thresholds — cost versus a human baseline, correction rate, security incidents — whose breach triggers a documented withdrawal proposal, making de-adoption a designed decision instead of a defeat.

Context

An agent runs in production doing work people used to do. Its economics and quality drift over time: a provider change degrades accuracy, scope creep multiplies cost, correction work quietly shifts back to humans. The team that built it evaluates it, and every incremental fix feels cheaper than admitting the deployment is not paying off.

Problem

Nobody plans the exit. Withdrawal reads as defeat, so teams escalate commitment instead: one more prompt fix, one more guardrail, while the agent costs more than the humans it replaced and correction rates climb. Because degradation is gradual, each week's numbers are individually defensible, and criteria negotiated in the middle of a failure are political — the people who would decide have already sunk their credibility into the deployment. The result is a fleet of half-working agents nobody dares to kill and no record of when the case for them stopped holding.

Forces

  • Escalation of commitment is well documented for software projects: sunk cost and personal attachment keep failing efforts alive long past the evidence.
  • Degradation is gradual and multi-axis — cost, quality, risk — so without pre-agreed bounds no single data point ever forces the question.
  • Criteria agreed before go-live are cheap and neutral; the same criteria proposed during a failure are contested by everyone invested in the deployment.

Example

A team ships a knowledge-base QA agent with agreed thresholds: cost per resolved query under 1.5 times the help desk baseline, correction rate under 50 percent, immediate halt on any data-leak incident, review after six months. A model provider change pushes the correction rate over the bound for two consecutive weeks; the on-call records the breach and files the withdrawal proposal. The review narrows the agent to the two document domains where corrections were rare — the descent path the criteria had designed in advance.

Diagram

Solution

Therefore:

Define the exit before the entry. Agree measurable bounds on three axes: economics (cost per case above a multiple of the human-equivalent baseline, such as 1.5x), quality (human-correction rate above a bound, such as half of outputs needing rework), and risk (incident classes that halt the agent immediately). Give every deployment an expiry date — a review point at which continued operation must be re-justified rather than assumed. Instrument the measurements from day one so the thresholds are evaluated by pipeline, not by advocacy; when a bound is crossed, record the fact and present it as the basis for a withdrawal proposal. The decision itself stays human and has three designed outcomes: keep with a re-baselined threshold, narrow the agent to the scope where it performs, or withdraw — executed as a staged descent of autonomy and scope, with salvage of what the deployment learned, rather than an overnight unplug.

What this pattern forbids. Thresholds must be agreed before go-live and cannot be renegotiated mid-breach: a crossed bound is recorded and forces a documented keep, narrow, or withdraw decision, a defined security incident halts the agent immediately, and no deployment runs without an expiry date for its next review.

The smaller patterns that complete this one —

  • usesCost Observability★★— Surface per-request, per-user, and per-feature cost and token consumption to operators in near-real-time.

And the patterns that stand alongside it, or against it —

  • complements[crawl-walk-run-automation-gating]
  • complementsKill Switch★— Provide an out-of-band control plane to halt running agent instances without redeploy.
  • conflicts-withPerma-Beta✕— Anti-pattern: ship the agent in 'beta' indefinitely so that quality regressions are someone else's problem.

Neighbourhood

Click any neighbour to follow the language. Scroll to zoom, drag to pan.

References

Provenance