Blog
AI Agents: One Success Is Not an Operations Test
AI agents can complete an instruction correctly once and take a different path on the next run. IBM Research uses an AppWorld test to show why an average success rate is insufficient: a ReAct agent with GPT-4.1 passed 77.4 percent of runs, but passed only 53.0 percent of tasks in all five repetitions.
Accuracy and consistency are separate metrics
The usual average rate tells you how often an agent succeeds overall. It does not tell you whether a workflow that passed today will remain reproducible tomorrow. IBM describes the difference in its example as a 24.4 percentage-point consistency gap.
The proposed Consistency Analyzer examines recorded trajectories and identifies decision points where small changes can flip the next action. It repeatedly samples individual steps under controlled conditions without rerunning the entire business process. According to IBM, guidelines generated from this analysis reduced the measured gap from 24.4 to 12.0 percentage points in the experiment without reducing average accuracy.
Optimisation needs a release process
AWS takes a complementary approach with Amazon Bedrock AgentCore. Production traces and a reward signal are used to recommend system-prompt changes. Teams can validate variants in offline batch evaluations and then in controlled online A/B tests before promoting a new configuration.
Both approaches expose the same operational mistake: a successful demo run is not evidence of a stable agent. An automatically optimised configuration should not become the new default immediately either. It needs a fixed test suite, comparative metrics and an explicit approval point.
What enterprises should measure
For a production-oriented pilot, measure at least four values separately:
- success rate across all runs,
- the share of tasks that pass every repetition,
- the type and location of divergent tool decisions,
- cost and runtime per repeated test.
Repeat critical workflows with the same instruction and a controlled environment. Examples include contract review, account reconciliation, ticket routing and changes to production systems. Define which variation is acceptable and when a human must take over.
The second run is therefore not redundant. It tests a different property: not whether your agent can complete the task, but whether its behaviour is reliable enough for operations.