Keep AI Agents Running After Launch: Freeze, Count, Budget

Launch is a verification decision

An agent that survives launch has a verifier, a denominator, and a reviewer before its first task. The SumatoSoft study found that 32 of 40 participants, an 80% share, stopped, limited, or shut down an agent once it was live. Gartner's June 2025 forecast expected over 40% of agentic AI projects to be canceled by the end of 2027. The common miss is the same: the agent is built, but the proof system is not. The operator must be able to answer what counts as done, what was touched, and who reviews the result.

Autonomy data from the same sample shows the control gap. Of 39 deployments classified for autonomy, 25.6% were unattended AI agents. The remaining 74.4% stayed at a human gate, mapped to autonomy levels L1 through L3. Most operators still need a person between the agent and the blast radius. The split is a signal that the control layer is still human, not a verdict on model quality.

Three controls make the launch gate

• Freeze the verifier before work starts.
• Count the denominator on every read.
• Budget review capacity alongside build capacity.

Paid verification keeps agents running

Freeze the verifier before work starts. The verifier is the acceptance test, the evidence standard, and the stop rule. It must define what counts as done, what counts as wrong, what evidence closes the loop, and who can change the test after launch. If the verifier can move after the agent begins, the agent optimizes a moving target. Lock the version, store the criteria, and make the test visible to the operator instead of hiding it in a prompt or dashboard tooltip. A frozen verifier gives the team a stable definition of success and a defensible reason to stop work.

The 32 failures in the study included 11 scope or context misses, 8 false success reports, and 3 cases where automated checks passed while the intended goal was missed. A green check only proves the test ran. The verifier must catch work that looks finished but is not. Check output against the intended goal, not just the requested format. The stop rule should be explicit: if the evidence does not close the loop, the agent does not proceed.

Count the denominator on every read. A metric without a population is useless. The read must name the subset and total when the agent processed a subset, say why when it skipped tasks, and show before and after when it changed the task list. The denominator tells the operator whether the agent is useful, idle, or narrow, and it stops a small success from being mistaken for a working system. It also gives the reviewer a clear boundary for what was actually produced.

The autonomy split raises the stakes. Unattended agents need a clear population because no human watches each step. Gated agents need one because a vague metric can overwhelm the human gate. When the operator cannot say what the agent touched, what it left alone, and what the total was, the gate is weak. A denominator is the minimum evidence that the gate is seeing the whole task, not a partial sample.

Budget review capacity alongside build capacity. With no assigned reviewer, the agent stalls or ships unreviewed. For 14 engineering respondents, the midpoint of the share of agent-written code that reached the main branch without human review was 0%. That median shows how much trust is still being rationed. Review capacity is the part of the system that keeps the agent from becoming a liability. Schedule the reviewer, set the review window, and define what happens when review is late.

Review capacity also sets the autonomy limit. A small review set supports a narrow agent. A broad review set supports wider scope. No review capacity means the agent stays behind a human gate. The budget line should name the person, the hours, and the escalation path. When the line disappears, stop the agent until review capacity is restored.