Will Larson on Why Pure Agent Workflows Fail

Will Larson lays out a practical case against fully agent-driven automation, drawn from a real Dependabot security alert pipeline that kept misclassifying severity levels. His fix was to move filtering and flow control back into deterministic code and limit the agent to genuinely ambiguous tasks – like figuring out which team owns a piece of code via GitHub codeowners files. The three-phase pattern he describes (prototype with agents, refactor logic into code, scope agents down to the hard parts) is increasingly where experienced engineering teams are landing. It runs counter to the prevailing vendor pitch of end-to-end autonomous agents, but the reliability tradeoff is hard to argue with.