[{"data":1,"prerenderedAt":107},["ShallowReactive",2],{"resource-\u002Fresources\u002Fa-human-in-the-loop-that-actually-holds":3},{"id":4,"title":5,"author":6,"body":7,"date":97,"description":98,"extension":99,"meta":100,"navigation":101,"order":102,"path":103,"seo":104,"stem":105,"__hash__":106},"resources\u002Fresources\u002Fa-human-in-the-loop-that-actually-holds.md","A human-in-the-loop that actually holds","DevStride",{"type":8,"value":9,"toc":88},"minimark",[10,14,17,20,25,28,31,34,38,41,44,48,51,54,58,61,64,68,71,74],[11,12,13],"p",{},"Every AI-in-delivery vendor now says the same three words: human in the loop. It has become the reassurance you reach for when someone worries about handing judgment to a machine. Say it and the worry is supposed to go away.",[11,15,16],{},"It should not go away that easily, because most human-in-the-loop setups do not survive contact with a busy week. A loop where the person confirms whatever the agent already decided is not oversight. It is a slower way of doing what the agent wanted, with a human name attached for comfort. The rubber stamp is worse than no human at all, because it manufactures the appearance of judgment where none happened.",[11,18,19],{},"A loop that actually holds has three properties. None of them are about adding a human. They are about making the human's presence mean something.",[21,22,24],"h2",{"id":23},"the-human-is-deciding-not-confirming","The human is deciding, not confirming",[11,26,27],{},"The test is simple. In your current setup, could the person in the loop realistically say no?",[11,29,30],{},"If the only information they get is \"the agent did X, approve?\" then no is not a real option, because they have nothing to weigh it against. They approve because refusing would mean redoing the analysis themselves, which is the whole thing they were trying to avoid. That is confirmation wearing the costume of a decision.",[11,32,33],{},"A real decision needs the tradeoff on the table: what changes, what it affects, what it costs, and what the alternative was. When the agent surfaces the reasoning instead of burying it in a result, the human can actually exercise judgment. They can approve, adjust, or reject, and any of the three is a live outcome.",[21,35,37],{"id":36},"the-moments-of-judgment-are-named-in-advance","The moments of judgment are named in advance",[11,39,40],{},"Not every action needs a person. If you route all of them through a human, you have not built oversight, you have built a bottleneck, and bottlenecks get bypassed. The skill is deciding ahead of time which moments genuinely require judgment and which an agent can carry on its own.",[11,42,43],{},"A useful default: an agent can move freely inside a boundary and must stop at the edges of it. Reading the plan, drafting a breakdown, sequencing dependencies, preparing a change, these are inside the boundary. Applying a change to shared work, shipping to production, altering scope or commitments, these are edges. Name the edges explicitly and the loop stops being a vague promise to \"stay involved\" and becomes a specific contract about where humans decide.",[21,45,47],{"id":46},"the-record-makes-the-judgment-inspectable","The record makes the judgment inspectable",[11,49,50],{},"A decision no one can trace is only half a decision. When the human's approval and the agent's proposal both land on one shared record, with names and an order, the loop becomes auditable after the fact. You can answer why something happened, who approved it, and what it was weighed against.",[11,52,53],{},"The trail matters more with AI in the mix, not less. The faster the contributor, the more important it is that its work leaves a record a person actually signed. Attribution is not bureaucracy here. It is the thing that lets you trust speed.",[21,55,57],{"id":56},"what-this-looks-like-in-practice","What this looks like in practice",[11,59,60],{},"The DevStride delivery loop draws the boundary this way. An agent does the legwork of planning and building: it drafts the plan, breaks down the work, and prepares changes. Every change is a proposal until a person approves it. Code an agent writes lands in staging and reaches production only on a human sign-off. And every action, human or agent, is recorded with a name.",[11,62,63],{},"The result is not less AI. It is more, used harder, because the boundary is clear enough to trust. When people know exactly where their judgment is required, they stop hovering over the parts that do not need them, and the speed shows up.",[21,65,67],{"id":66},"the-honest-version","The honest version",[11,69,70],{},"You cannot buy a human-in-the-loop. You design one. It holds when the person can say no, when the moments that need them are named, and when the whole thing is on the record. It fails when \"human in the loop\" is a line on a slide and the human is a formality.",[11,72,73],{},"Speed and judgment are not a tradeoff you have to make. They are a boundary you have to draw, and then enforce in the workflow so it holds on the hard weeks too.",[11,75,76,77,82,83,87],{},"One of four principles in ",[78,79,81],"a",{"href":80},"\u002Fresources\u002Fthe-ai-era-delivery-playbook","the AI-era delivery playbook",". The enforcement mechanism behind it is ",[78,84,86],{"href":85},"\u002Fresources\u002Fguardrails-for-ai-in-delivery","propose, approve, record",".",{"title":89,"searchDepth":90,"depth":90,"links":91},"",2,[92,93,94,95,96],{"id":23,"depth":90,"text":24},{"id":36,"depth":90,"text":37},{"id":46,"depth":90,"text":47},{"id":56,"depth":90,"text":57},{"id":66,"depth":90,"text":67},"2026-07-29","\"Human in the loop\" has become a slogan people say and do not build. A loop where the person rubber-stamps whatever the agent produced is theater with an extra step. Real oversight means the human is making a decision, not confirming one. Here is how to draw that line honestly.","md",{},true,6,"\u002Fresources\u002Fa-human-in-the-loop-that-actually-holds",{"title":5,"description":98},"resources\u002Fa-human-in-the-loop-that-actually-holds","OlK_mnU5CD5_4pbf39TJLVa4VoPiat6oEkVkQzCpz8w",1785519825858]