AI Agent Shipped 3 Pull Requests in 75 Minutes, But Needed 12 Corrections Along the Way

A developer used a multi-model AI pipeline — Claude for planning, DeepSeek as orchestrator, and Codex for implementation — to submit three pull requests to an open-source Microsoft repository in roughly 75 minutes of active work. The session produced about 3,500 lines of code across an MCP client web part, an Azure AI Agent chat, and an M365 Copilot Agent chat, all of which passed automated validation. Out of 30 messages sent during the session, 12 were corrections, representing a 40% steering rate despite the overall success. The developer found that most errors clustered around two phases: establishing the pipeline workflow and fixing validation warnings, both triggered by the agent attempting to handle tasks itself rather than delegating. Notably, three corrections stemmed not from reasoning failures but from the agent ignoring context it already had, such as using an outdated framework version or the wrong SDK despite being told otherwise.