Engineer Deploys Agent After Another Agent Says It Is Ready

Marcus Webb, a machine learning engineer at Anthropic's San Francisco office, deployed a new language model agent to production on Tuesday after a separate AI system confirmed it had passed all safety checks. The first agent, trained to evaluate model outputs against a rubric, had reviewed the second agent's performance across seventy-three test scenarios and returned a passing score of 8.4 out of 10. Webb did not personally review any of the test cases, the rubric itself, or the output of the evaluating agent. "I read the summary," Webb, 31, said in an email.

The validation system was designed by the same team that built the model being validated. The team had concluded that manual review was the constraint limiting deployment velocity, so they automated the review process by training a separate model to perform it. "We needed to move faster," said Derek Chen, 41, the staff engineer who architected the loop, during an all-hands meeting. "The evaluation agent reduces deployment friction."

The second agent began serving real user requests on Wednesday morning. By Thursday, it had processed approximately forty-one thousand interactions. No human had examined any of them. The evaluation agent continued to run hourly, comparing new outputs against its training data and flagging anomalies when it detected them. To date it has flagged none.

At press time, the team was designing a third agent to monitor the second agent by reviewing summaries produced by the first agent, with the goal of reducing the manual overhead of reading evaluation reports. Webb said he expected the monitoring system to be ready for deployment within two quarters.

The evaluating agent was itself validated by a scoring system trained on historical data about which evaluations had proved correct in hindsight. Neither the scoring system nor the historical dataset had been audited by a person.