Where the agent sits in the workflow

An agent is a step in someone's week, and attention is the budget it spends.

The idea

A working agent is not judged on its outputs. It is judged on what somebody's week looks like afterwards. So the design questions are about placement: what fires it, how often it is allowed to speak, what it hands over and when, and what it does on the days it has nothing useful to say.

Two dials decide most of it. A confidence threshold chooses what it handles by itself against what it passes up, and moving it trades errors for escalations. A cadence cap limits how many times a day it may interrupt, which protects the one resource that does not scale.

The failure that ends most internal agents is not a wrong answer. It is being noisy enough that people stop reading it, at which point the correct outputs stop counting too.

Confidence threshold 75 Cadence cap 4

How it works

The routing is three branches and a budget. Everything interesting is in where the two numbers are set.

def route(item, threshold, budget_left):
    if item.confidence >= threshold:
        return "handle" if budget_left else "hold"     # cap protects attention
    if item.confidence >= threshold - 25:
        return "escalate" if budget_left else "hold"   # ask, do not guess
    return "quiet"                                     # nothing useful to say

def run_day(items, threshold, cap, trigger):
    budget = cap
    routed = []
    for item in sorted(items, key=lambda i: -i.confidence):
        action = route(item, threshold, budget)
        if action in ("handle", "escalate"):
            budget -= 1
        routed.append((item, action))
    # a digest is one interruption for the whole day; per-event is one each
    speaks = [r for r in routed if r[1] in ("handle", "escalate")]
    interruptions = (1 if speaks else 0) if trigger == "digest" else len(speaks)
    return routed, interruptions

Sorting by confidence before spending the budget matters: the cap should be spent on what the agent is surest about, not on whatever arrived first.

Cost

Choice	Time saved	Attention spent	Risk
Daily digest	High	One interruption a day	Latency: urgent things wait for the digest
On every event	High	One per item	Notification fatigue, then it gets muted
Only when asked	Low	None	Depends on someone remembering it exists
Lower the threshold	More handled	Same	More handled wrongly, and cleanup costs double

Cleanup is the number people forget to model. An item the agent gets wrong does not cost zero, it costs more than doing it by hand, because someone has to notice, undo, and redo it.

Watch out for

No quiet default. An agent that reports every day, including days when nothing happened, teaches people that its messages are ignorable. Saying nothing has to be a valid outcome, or the useful messages arrive in a channel nobody reads.
A cadence cap that silently drops work. Capping interruptions is right. Discarding the items that did not fit is not: they need to be held and carried forward, and the count of what is waiting has to be visible somewhere, or the cap turns into invisible data loss.
Escalating without the evidence. An escalation that says "this one needs a human" costs more than doing it, because the human starts from nothing. Hand over what it found, what it is unsure about, and what it would have done.
Measuring outputs instead of the week. "Handled 200 tickets" says nothing about whether the queue got shorter, since the same person may now spend the morning checking 200 outputs. Measure the time the work actually took end to end.
One threshold for every kind of item. Drafting an internal summary and closing a customer ticket do not deserve the same confidence bar. A single global threshold forces the safe case to be over-cautious and the risky one to be reckless.
Ignoring when it runs. An agent that posts its digest at 3am has its work read at the bottom of a backlog. Placement in the day is as much of a design decision as placement in the workflow.

Worked example

A marketing lead spends three hours every Monday assembling a performance digest from three tools. The agent's job is that assembly.

Set it as a daily digest at 08:00 with a cadence cap of one, because the whole point is a single artifact. Set the confidence threshold high, so anything it is unsure about is escalated inside that same digest rather than asserted. On a day where nothing moved, it says so in one line and stops.

Now try the panel with the threshold at 55 and the trigger set to every event. More items get handled, the tally of time saved goes up, and so does the count of items handled wrongly, each costing double to undo. The interruption count climbs from one a day to a dozen. Both configurations "work". Only one of them is still being used in a month.

Check yourself

Your triage agent handles 80% of tickets correctly and escalates the rest. Adoption is falling and people say it is "noisy". What do you change first?