Quick answerGraduate a task category to full autonomy, no human escalation path at all, only after it clears four bars independently of the agent's overall launch status: a large enough completed volume to estimate its true error rate with statistical confidence, a sustained error rate below a pre-agreed ceiling with no upward drift over the measurement window, a defined and bounded worst-case outcome if the agent gets it wrong, and an explicit sign-off from whoever owns the risk for that specific task category, not a blanket approval for the agent as a whole. Most agents should never graduate every task category at once; the bar is set and cleared one category at a time.
This is not the same milestone as launching the agent
Getting an agent from pilot to general availability, covered elsewhere in this series, is a decision about whether the agent overall is ready to serve real customer volume with an acceptable baseline of human oversight in place. Task-category graduation is a decision that happens continuously after that, often over many months, about whether a specific narrow slice of what the agent does has proven itself well enough to remove the human escalation path for that slice specifically. An agent can be fully launched and still have zero task categories operating with no escalation path at all, which is the normal, safe default, not a sign the rollout stalled.
Why category-by-category, not agent-wide
Task categories inside one agent vary enormously in downside risk and in how measurable their error rate actually is. "Reset a password" and "approve a five-figure refund" might both run through the same agent, but they do not belong on the same autonomy timeline. Graduating the agent as a whole to full autonomy, rather than category by category, means the riskiest task in the agent's repertoire sets the ceiling for how fast the safest task can be trusted, which either slows the safe categories down for no reason or pushes the risky ones forward before they have earned it.
The four-part bar
Volume: enough completed instances of that specific task, not the agent's total volume, to make the observed error rate statistically meaningful rather than a small sample that happened to look clean. Sustained low error rate: below a ceiling the business has agreed to in advance, measured over a rolling window long enough to catch drift, not a single good week. Bounded worst case: a wrong outcome in that task category has to be recoverable, reversible, or capped in size, since an unbounded worst case is a reason to keep a human in the loop regardless of how rarely it occurs. Owner sign-off: someone who actually owns the consequence of that task category going wrong, not a general engineering or product approval, has to explicitly approve removing the escalation path.
What graduation actually changes operationally
Removing the escalation path is not the same as removing oversight entirely. Most teams that do this well keep post-hoc sampling and monitoring in place for a graduated category, they simply stop requiring a human to approve before the action happens. The distinction matters: graduation removes the blocking checkpoint, not the visibility into what the agent is doing in that category after the fact.
Build in a way back down
A task category that graduates can also un-graduate if its error rate drifts upward or a new failure mode appears that the original measurement window never captured. Define the demotion trigger at the same time you define the graduation bar, using the same metrics, so pulling a category back under human review is a pre-agreed operational response rather than a reactive scramble after something goes visibly wrong.
FAQ
How many completed instances count as enough volume?
There is no universal number; it depends on the baseline error rate you are trying to detect. A task with a target error rate under one percent needs a much larger sample to measure confidently than one with a five percent ceiling. Work backward from the statistical confidence you need, not a round number that feels sufficient.
Can a task category graduate before the agent has been in general availability?
In principle yes, if a specific narrow task has enough pilot-stage volume and a bounded worst case, but in practice most companies wait until GA simply because pilot volume is rarely large enough to clear the statistical bar for any task category.
What is the most common mistake teams make here?
Treating graduation as a one-time decision instead of an ongoing measurement. The most common failure is graduating a category correctly, then never re-checking its error rate against the original ceiling as volume and edge cases grow.