AI agents can investigate problems, use external tools, and coordinate with other agents to complete complex tasks. But when an agent makes an incorrect assumption or generates a hallucination, subsequent actions may compound the error. Reviewing every step manually can be costly and difficult to scale.
Microsoft proposes an AI system that introduces a separate “critic” into an agent’s workflow. The critic reviews reports generated during tool use, identifies potential errors, and sends feedback to the original agent so it can correct its work before the workflow continues.

How agents check and correct their work
The system divides the workflow among an assistant, a user proxy, and a critic. The assistant recommends tools to investigate an event, while the user proxy executes the recommended tools and returns their results. The assistant then generates a report describing the tool response and sends it to the user proxy.
In this agentic setup, the proposed system introduces a second, separate conversation in which the user proxy forwards the report to the critic. The critic checks the report for errors, such as skipped investigation steps or an inappropriate tool recommendation, and suggests potential corrective actions. The user proxy then relays the critic’s findings to the assistant, which can revise its analysis and conclusions or recommend additional tool calls.
This tool-call-and-analysis loop can repeat until the critic determines that the report is free of errors. Notably, the filing specifies that the assistant is “naive to” the separate conversation between the user proxy and the critic. This keeps the review process separate from the assistant’s original interaction with the user proxy, allowing detected errors to be addressed as the workflow progresses rather than only after the task is complete.
Making autonomous workflows more reliable
The proposed approach addresses a central challenge in autonomous AI systems and agentic workflows. An agent can execute a sequence of actions successfully yet still reach an incorrect conclusion. Agent behavior can change as models, tools, and operating conditions evolve. If a flawed conclusion influences subsequent tool calls, errors can propagate throughout the workflow before the task is complete. Introducing a critic at intermediate stages could help identify problems earlier, reduce the need for manual review, and improve the reliability of complex, multi-agent tasks.
Cybersecurity is a particularly relevant application described in the filing. Security operations centers (SOCs) handle alerts involving suspicious emails, URLs, file attachments, and other potential threats. Microsoft’s proposed architecture allows a lead investigation agent to delegate tasks to specialized agents, each with its own assistant, user proxy, and critic. For example, separate agents could investigate a suspicious email’s attachment and URL, with their reports reviewed before being passed back up the investigation chain, streamlining the process.
The approach also aligns with Microsoft’s broader investment in agentic AI. In March 2025, Microsoft announced new agents for Security Copilot to automate tasks such as phishing triage, data-security alert review, and threat intelligence analysis. Two months later, the company also introduced multi-agent orchestration in Copilot Studio, enabling agents to collaborate and divide complex tasks according to their respective capabilities or expertise. More recently, Microsoft has emphasized agent observability, including tracing agent actions, evaluating quality and safety, monitoring behavior, and identifying opportunities for improvement. These initiatives reflect a growing recognition of the need to make increasingly autonomous AI systems more reliable, secure, and accountable as the industry moves toward a more agentic ecosystem.
U.S. Pat. App. Pub. No. 2026/0288566, titled “Artificial Intelligence-Automated Review of Agentic Flows,” was filed on March 21, 2025, and was published on September 24, 2026. The listed inventors are Duc Minh Le, Neelanjan Hector Jacob, Shane Anil Pereira, and Efim Sokolov.
