For most of the history of enterprise automation, “should a machine do this?” was essentially a binary choice — automate it or don’t. What’s changed in the last couple of years is that the question has become more granular: not whether AI is involved, but how much it drives, and under what conditions it earns the right to work without a human watching over its shoulder.
That shift in how we frame the question turns out to matter a lot in practice.
The structure underneath the workflow
A useful way to think about any regulated or policy-driven process is as two overlapping graphs. The first is the end-to-end workflow — a directed acyclic graph (a map of all the steps, where each step flows only forward, never back) showing how a piece of work moves from start to finish. The second is what happens inside each step: a smaller, cyclic loop where an agent works, checks its output, gets feedback, and iterates until the output is good enough to move on.
The principle that follows from this is worth keeping: keep the outer process deterministic, let the inner loop be adaptive. The overall workflow follows policy and doesn’t skip steps. The agent inside each step is allowed to reason, revise, and try again. That containment is what makes the adaptive part safe.
Four questions worth asking before handing a node to an agent
Not every step in a workflow is ready to run unsupervised. A practical set of diagnostic questions helps separate the candidates from the ones that still need a human:
1. Can the output be checked against policy? If there’s no clear standard for “correct,” there’s no way to know when the agent is done — or wrong.
2. Can it run in a bounded loop? Unbounded iteration is a risk, not a feature. Steps that could spin indefinitely without a natural exit condition are poor candidates for unsupervised operation.
3. Can it be traced? Auditability — the ability to reconstruct what the agent did and why — is a hard requirement in any regulated environment, and a strong signal of overall process health in any environment.
4. Can it fall back to a human? The ability to escalate when the agent hits an edge case is what makes the difference between fragile automation and resilient automation.
A “yes” on all four is a reasonable signal that a step is worth automating further. Decision nodes — the places where policy is actually interpreted rather than just applied — tend to fail at least one of these and appropriately remain human-driven longer.
Trust is earned through measurement, not time
The mechanism for moving a step toward full autonomy is straightforward in concept, though harder in practice. It has three phases.
In the first phase, humans verify everything the agent produces. The goal here isn’t efficiency — it’s calibration. This is when teams build the datasets and work with domain experts to define what “good” actually looks like in concrete terms. Without that definition, there’s no exit condition.
Once measured confidence meets a policy-defined threshold, the team moves to sampling: humans review a slice of output, and systematic evaluation via runtime logs covers the rest. The key word is measured — this is statistical confidence derived from representative datasets and repeated evaluation series, not a gut feeling about the model.
The third phase is exception-only: agents handle the bulk of the work unsupervised, humans step in for outliers. Crucially, the work of refining datasets and sharpening the definition of “good” continues. The confidence threshold isn’t a finish line; it’s a floor that gets raised.
The bottleneck shifts
One consequence of this approach that’s easy to underestimate: accelerating one part of a workflow with agents doesn’t make the overall workflow faster — it shifts the bottleneck. Faster code generation means design review or security scanning becomes the new constraint. Process visibility and metrics matter more than the tooling used to author the agents themselves.
This is a familiar pattern from any capacity-planning exercise, but it catches teams off guard when AI is involved because the throughput gains can be dramatic enough to expose downstream constraints that were previously invisible.
What this means for how we build
The framing here — DAG structure, evals as the confidence mechanism, a trust ramp built on measurement — is less about any specific tool and more about treating AI-driven work the way we treat any probabilistic system: with explicit quality criteria, ongoing measurement, and a clear escalation path. Agents are neither a magic accelerant nor a liability to be minimised; they’re a component in a system, and the system’s design determines whether they’re trustworthy.
As more workflows in more industries start operating this way, the teams that will do it well are the ones that approach autonomy as something earned through evidence — one step, one dataset, one threshold at a time.
Text summarized and optimized using Anthropic’s models and reviewed by a human.