The skills that matter when agents act alone on routine work

Depth 3 is where an agent stops asking. It executes routine work within bounds, and nothing reaches a person unless it falls outside them. The skills that matter then are sampling what was allowed, spotting drift, defining what counts as routine and keeping sources current. Most teams train for the held queue and neglect the work nobody sees.

What changes when an agent works alone?

At depth 2, every action passes a person: the agent drafts and someone sends. At depth 3, the agent has earned, on evidence from the business's own cases, the authority to carry out routine work itself. A finance agent posts recurring accruals below a threshold. A payroll agent processes standard pay changes. A support agent issues refunds within the business's limit.

That is the point of depth 3. It frees people from volume. But it also means a category of work no longer has a person looking at each item. If something goes wrong inside the bounds, nobody sees it in the moment. The record files every allowed action, so it can be reconstructed, but someone has to look.

Which skills matter most?

Sampling allowed work

Held items demand attention because they sit in a queue. Allowed items do not. The first skill is looking at a sample of what the agent did alone, regularly, and checking it as carefully as a held item. Good sampling is deliberate: some random, some targeted at the edges of what counts as routine, some at items that were close to a limit.

Spotting drift

An agent's behaviour can change gradually as inputs change: new suppliers, new product lines, new customer types. Its certificate lapses after 90 days, and recertification against fresh cases will catch drift, but only at that point. People who sample and read the record can spot it earlier. Attest's stable gate checks that the agent gives the same answer to the same question across runs, but only when a run happens; the people around the agent should be watching for drift in between.

Defining routine

"Routine" is a business decision. Which journals, which pay changes, which refunds are routine enough for the agent to handle alone? The people who set that boundary need to understand the work deeply, which is where subject matter experts earn their place. A boundary set too wide lets non-routine work through without review; set too narrow, it floods the held queue.

Keeping sources current

An agent acting alone acts on its sources. If a policy changes, a rate table is updated or a supplier list goes stale, the agent's routine work goes wrong silently. Someone must own keeping those sources current and check that the agent is reading the latest version.

How does this look in specific roles?

For a financial controller, the agent posting routine journals means the controller's main contribution shifts from reviewing every journal to sampling posted ones, setting thresholds deliberately and checking that the close still makes sense as a whole. The skills that matter most for financial controllers go further into this.

For a payroll administrator, routine pay changes processed by the agent mean the administrator's attention moves to the exceptions and to sampling standard changes for errors that would otherwise surface only when an employee complains. See the skills that matter most for payroll administrators.

In both, the job is less about deciding and more about watching well.

Who should do the watching?

Not only the most senior person, and not only the most junior. Sampling allowed work needs enough experience to recognise a subtle error, and enough time to do it properly. Many teams split it: experienced reviewers sample the edge cases, newer ones sample at random under guidance. The approach in how to split review work between junior and senior staff applies to sampling as much as to held items.

Managers matter here too. Whether sampling actually happens week to week depends on whether managers protect time for it, which is one reason middle managers decide whether agents work.

What is the trade-off?

Sampling allowed work costs time that depth 3 was meant to save. If people re-review everything the agent does, there was no point granting the depth. Too little sampling, and errors accumulate unseen.

There is also a trust question. Heavy sampling can signal that the team does not trust the agent's certification; light sampling can signal complacency.

The answer is proportion. Sample enough to detect drift and boundary errors, weighted towards the edges of what counts as routine. Increase sampling when inputs change or a recertification is approaching; reduce it when the agent's record is clean and stable. Treat sampling as part of the agent's governance, not as a vote of confidence or distrust.

What should you do differently?

Add sampling of allowed work to the weekly rhythm, with a named owner and a defined method. Teach reviewers what drift looks like. Write down what counts as routine and who decided. Assign someone to keep each source current. And train for these skills explicitly, because they do not come from learning to approve held items. The guide to AI agent skills by role sets out what each role needs.

Where the Academy fits

Pryme Intelligence Academy's Operator track includes modules on reading the workspace, keeping knowledge current and running the rhythm, which cover sampling and source upkeep as well as review. Module tests are set in the learner's own industry and role, so a payroll administrator's scenarios differ from a controller's.