Most decisions do not die from a definitive “no.” They die from a quieter thing: the cost of finding out. A household move is a good example. The paperwork, the utility transfers, the lease timing, the school district research, the removal quotes — none of it is conceptually hard, but the aggregate is roughly twenty hours of unglamorous investigation. That number is large enough that many people simply do not start. The move stays a vague idea rather than a real option.
This is worth sitting with. A possibility you cannot afford to investigate is barely a choice at all.
What Changes When Investigation Gets Cheaper
There is a tendency to measure AI agent progress in terms of speed: how much faster does an existing task get done? That is a real gain, but it may be the smaller one. The more consequential shift is that a more capable agent changes which tasks get attempted at all — it brings choices back into range that were previously too expensive to pursue.
Running a simulated household move through an agent makes this concrete. The question given to the agent was not “help us move” but something structurally different: “help us understand whether moving would improve our lives.” That reframing matters. The first is a logistics problem. The second is an investigation, and it needs to happen before any decision is made.
Recent agent benchmarks reflect this shift in capability. AutomationBench — a measure of how far an agent can drive a multi-step real-world task — has roughly doubled in completion rate year over year for leading models, though the absolute number (around 41 percent) still means oversight is required throughout. These are partial-credit scores, not perfect completions: a capable agent gets most of the way through most tasks, but it still surfaces wrong assumptions, follows weak leads, and occasionally carries a flawed premise a long way before anything flags it.
That last point matters for how to use an agent on a consequential investigation. The most common mistake is to hand it a plan and ask it to execute; a better move is to hand it an uncertainty and ask it to investigate. The difference is subtle but important — an agent given a plan will optimize within the plan’s assumptions; an agent given an open question will surface whether the assumptions hold.
Framing the Assignment
The 23-card guide that emerged from the simulated move is essentially a checklist for scoping an investigation correctly. Each card addresses something the agent needs to know before it can do useful work: the goal, the constraints (budget, timing, geography), where its authority ends, and what questions it should surface back rather than decide unilaterally.
A lighter framing like this — specifying the destination and the guardrails, then letting the model find its own method — tends to outperform heavily scripted instructions. The agent does not need to be told to search in a particular order; it needs to know what a good result looks like. Habits formed when agents were weaker — breaking every task into tiny explicit steps, pre-scripting the sequence — can actually prevent delegation of the real investigative work.
One technique worth keeping from the simulated move: after the agent returns options, ask it to investigate what would make its own recommendation wrong. This adversarial self-check surfaces the evidence it weighted most heavily and flags where that evidence is thin — vendor claims, repeated sources, or a single article doing a lot of work.
Investigation as Access
There is an older version of this pattern. Research on financial aid applications found that combined application assistance plus information increased college enrollment by roughly eight percentage points, while information alone produced no measurable effect. The difference was someone who could push through the awkward interface — the forms, the follow-ups, the missing information — long enough for the help to land. Thoroughness has historically been unequal, available to those with assistants or time.
Capable agents extend that kind of help more broadly. Not evenly, not perfectly, but meaningfully. The household move is a stand-in for a whole class of decisions: the supplier switch nobody got around to pricing, the refinancing that seemed complicated, the career change that felt like too much research. Each one had a twenty-hour question mark in front of it.
As agent completion rates continue to improve, the practical question is less “what can an agent do?” and more “what did we quietly stop trying to figure out?” That second list may turn out to be the more interesting one.
Text summarized and optimized using Anthropic’s models and reviewed by a human.