- AWS's sales operations team says its agents can give fluent, wrong answers while monitoring stays green, so it tests answer correctness before and after every release.
- In one pre-production test, AWS says mismatched releases cut a pass rate from 78 percent to 38 percent. That is one test, not a general rate.
- Leaders should ask who defines a correct answer, whether data and prompt changes are tested, and how live answers are scored.
An agent can be online, fast and wrong
Most dashboards tell you whether a system responded. They do not tell you whether it responded correctly. With AI agents, that gap is where trust gets lost. The lesson here is simple: uptime shows an agent is talking, and only testing shows whether it is right.
AWS made the point in a post published October 16. Its Sales, Marketing, and Global Services Operations team runs a multi-agent system for AWS field sales. Sellers ask business questions in plain English. Specialized sub-agents handle territory performance, customer intelligence, and text-to-SQL queries over company data. Answers are scoped to each user's territory and role.
Before launch, the team evaluated its Customer Insights Agent and found wrong answers. AWS says monitoring stayed green through the different runs. Those green lights only confirmed that the system had replied. They said nothing about whether the reply was correct. AWS notes that a seller who passes an overstated number to a customer risks that customer's confidence.
Why the usual alarms stay quiet
Traditional software gives the same output for the same input. Agents do not. AWS notes they can follow different reasoning paths for identical inputs. They also make a chain of decisions: which sub-agent to call, which tool to use, how to combine the results. One weak step can quietly corrupt everything after it.
Think of a factory that counts boxes shipped but does not open them. The count looks healthy while defects pile up. AWS describes the same pattern: latency stays low and error rates stay flat while answer quality erodes.
The risk also moves without any release. AWS says upstream pipelines refresh the data every week. Releases can go wrong too. Some releases bundle a data update, a tool update and a prompt edit. If one piece arrives late, the system runs on mismatched parts until the rest catch up.
AWS measured the effect in one pre-production test. The mismatch window alone pushed passing answers down from 78 percent to 38 percent.
How AWS checks for it
AWS describes nine stages, from first commit to live production. They include testing each change, a gate before release, a canary check, adversarial testing, drift monitoring in live traffic, and a feedback loop that turns failures into new tests.
AWS says the framework covers five dimensions throughout the lifecycle: safety, quality, performance, cost and operational control. It also applies at three levels: single tools, each sub-agent, and the full multi-agent flow. Many teams stop at the agent level. AWS argues that a multi-agent system adds a failure surface neither of the other levels can see alone.
The base of it is a golden set. This is a curated list of real questions with known correct answers. AWS builds it from sampled production queries, past failures and deliberately constructed edge cases. It calls the set a living document. For the agent level, it sorts questions into four types and revisits them monthly.
AWS highlights two scoring approaches. Deterministic checks compare facts to a known answer. For judgment calls, an LLM-as-judge approach has a second model score the response against a written rubric. AWS says the first gives precision on facts and the second gives breadth on reasoning.
Read it with care
This is a vendor describing its own internal system, built on its own product, AgentCore Evaluations. The 78-to-38 figure comes from one pre-production test. It is not a rate for agents in general. The portion of the post reviewed here does not give the cost of running this testing.
The pattern still carries over to any vendor. Define what a correct answer is. Test it against known cases. Re-test whenever data, prompts, tools or models change.
Questions to put to your team
First, who has written down what a correct answer looks like for each agent, and is it a set of test cases with known answers? Second, do data and prompt changes trigger the same tests as code changes? AWS's weekly data refresh shows quality can move with no release at all.
Third, can linked changes to data, tools and prompts land apart, and who checks the system in between? Fourth, which of your dashboards measure whether an answer was right, rather than whether one arrived? Fifth, is a sample of live answers scored on a schedule, with the low scorers fed back into testing?
A green dashboard says the agent is answering. It does not say the agent is right.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: AWS.
