One agent watches the run, and a failure opens a conversation to fix it

A codified workflow used to run each of its agent steps as an isolated model call with its own transcript, and it handled a failed code step with a policy chosen back when the workflow was written: retry the same deterministic code on the same inputs, or replace code you had approved with an agent improvising. Both of those are gone.

Consecutive code steps now run as a chain with no model in the loop at all, so a fully codified workflow makes zero model calls and never builds an agent. One agent sits behind the whole run and is engaged at exactly two moments: a step you declared needs judgment, and a code step that failed. It shares one transcript across the run, so the agent judging step seven remembers what it saw when step three broke.

When it gets a step past a failure by revising that step's inputs, the run carries on and you get told. Nothing broke, but the workflow just proved it cannot execute as written, and that is worth knowing.

A failure it cannot fix opens a repair interview rather than a machine verdict. The interview sees the failing step in full, the code, the run's own history and your accepted examples side by side, makes the smallest change it can, and you confirm it as the next version. When the failure is not a bug, it says so and recommends handing the work back to an agent, and you decide. Your inbox links the failure straight into that conversation.

A workflow never quietly un-automates itself any more. The switch that let it do that has been removed.