There is a comfortable story we tell ourselves about agents and the build: that the human stays "in the loop", reviewing everything, and so accountability is preserved. It is a story that survives precisely because it is rarely examined. When agents generate a feature branch in minutes and a senior engineer is nominally responsible for ten such branches a day, the loop is not a control. It is a rubber stamp with a person attached. The uncomfortable truth is that as agents absorb more of the build, keeping a human in every loop does not preserve accountability. It quietly destroys it, by spreading attention so thin that nobody is actually deciding anything.
This is the deeper consequence of the Acceptance Gap. Once generation is abundant, the constraint moves to acceptance, and the scarce human work concentrates into a small number of genuinely decisive moments: framing the intent, judging whether the result fits, owning the trade-off, and accepting the residual risk. Most teams get this wrong in one of two directions. They either keep humans in every loop and forfeit the agent's leverage, or they remove humans entirely and forfeit the accountability that someone, eventually, will be asked to produce. The discipline that survives both failures is to design the acceptance loop deliberately: to decide which decisions require a human, at what altitude, with what evidence.
The evidence that the build is moving, and that trust hasn't
The shift is not speculative. Gartner projects that around 70% of enterprises will deploy agentic AI within IT infrastructure operations by 2029, up from under 5% in 2025. The build is genuinely migrating toward agents. What has not migrated is trust, and for defensible reasons. In the 2024 DORA research, AI adoption raised individual productivity and job satisfaction yet was associated with a measurable degradation in delivery stability and throughput, with one analysis of the data putting the stability cost at roughly 7.2% and the throughput cost at around 1.5% for a notional increase in adoption. The mechanism is unglamorous: larger AI-generated changesets carry more risk, and the old fundamentals of small batches and robust testing still hold. In the same survey, more than 39% of respondents reported little or no trust in AI-generated code.
Reviewers are voting with their judgement. A study spanning roughly eight million pull requests found only about 32.7% of AI-generated code was accepted, with reviewers exchanging nearly 12% more rounds on AI-authored changes than on human-authored ones. Read carefully, that rejection rate is not a failure of the agents. It is the acceptance loop doing its job, expensively and indiscriminately, on every change rather than the ones that warrant it.
Calibration, not maximisation
The instinct to put a human on everything misreads the goal. Microsoft Research's 2024 synthesis on appropriate reliance frames the target precisely: the aim is calibrated reliance, avoiding both overreliance, where wrong output is accepted uncritically, and underreliance, where correct output is ignored. Neither maximal human involvement nor minimal involvement is the objective. The objective is reliance proportional to evidence and risk.
The question is never "is there a human in the loop?" It is "is the right human at the right altitude, with evidence good enough to make the trade-off they are accountable for?"
Calibration is a design problem because the signals that help humans calibrate can be engineered. A 2024 ACM FAccT study with 404 participants found that first-person expressions of model uncertainty, such as "I'm not sure, but...", reduced overreliance on LLM answers. The caveat matters as much as the finding: the precise wording carries the effect, and a poorly designed explanation can backfire and increase overreliance. Evidence presented to an acceptance moment is not neutral. It is part of the control, and it can be built well or badly.
Autonomy is a dial, not a property of the agent
The most useful conceptual move comes from the Levels of Autonomy for AI Agents framework (Feng, McDonald and Zhang, 2025, at the Knight First Amendment Institute). It defines five escalating levels by the human's role, from operator to collaborator, consultant, approver and observer, and makes the point that designers most often miss: autonomy is a design decision that need not be coupled to capability. A highly capable agent can be deliberately constrained to elicit human feedback. You do not accept maximal autonomy because the agent is able to operate there; you choose a level because the decision warrants it.
This reframes the acceptance loop as a set of altitude choices rather than a single on/off switch. Routine, reversible, low-blast-radius work can sit at observer or approver level. Decisions that are hard to reverse, that touch trust boundaries, or that encode a strategic trade-off should pull the human up to collaborator or consultant, where they are framing and judging rather than rubber-stamping. The error most teams make is applying one altitude everywhere.
Concentrating judgement where it pays
Where, concretely, should the human moments sit? Practitioner guidance, including Qodo's 2025 State of AI Code Quality work, is consistent that AI-generated code carries materially higher defect and logic-error rates and that review should not be skipped for code touching authentication, payments or security. That is the design principle stated plainly: concentrate human judgement on high-risk surfaces and let it thin elsewhere. McKinsey's State of AI 2025 finds that high performers are nearly three times more likely to redesign workflows around AI rather than layer it on, and far more likely to have defined human-in-the-loop validation, at 65% versus 23%. The differentiator is not whether humans are involved. It is whether the involvement is designed.
Why the loop must be real, not symbolic
Regulation is converging on the same discipline and raising the stakes. EU AI Act Article 14 requires high-risk systems to be designed so they can be effectively overseen by natural persons who understand the system's limits, stay alert to automation bias, interpret output correctly, and can decide to disregard it, with oversight commensurate with risk, autonomy and context. Those obligations become enforceable from 2 August 2026, and deployers must assign oversight to specific, appropriately trained individuals. The warning sits alongside the requirement: analysis of NATO, US DoD and EU policy documents finds oversight in practice is often more symbolic than substantive, undone by compressed decision cycles, automation bias and vague definitions of human control. A loop that exists on paper but cannot actually intervene is the failure mode the regulation is trying to prevent, and it is the same failure mode that destroys accountability inside a delivery team.
There is a harder caution, too. A 2025 paper argues that fully autonomous agents should not be developed at all, because as autonomy rises, human control and accountability fall and safety risk scales with them. Whatever one's view on the absolute claim, the directional logic is sound and it is the logic of the acceptance loop: deliberate caps on autonomy are how leverage and accountability survive together rather than trading off against each other.
Designing the loop
A well-designed acceptance loop answers four questions for each class of change. Which decisions require a human, rather than all of them? At what altitude does that human sit, framing and judging rather than checking line by line? With what evidence, including calibrated uncertainty signals built to inform rather than to reassure? And who, specifically and by name, owns the trade-off and accepts the residual risk? Get those right and the agent's leverage is real because the human is not gating everything, while accountability is real because someone decisive is standing at each moment that matters.
The deeper move is to treat acceptance as a first-class part of delivery design, not an afterthought bolted onto code review. If acceptance is the constraint, then the loop that governs it is where your delivery system either compounds or quietly breaks. That is the argument we develop in The Acceptance Gap, and it is the right next read for anyone designing where people must stay in the system as the build moves on without them.