Those three all describe what the agent does. The dimension that decides adoption in the orgs I have seen is what happens after the agent stops: the rework a person does to make the output usable. That cost rarely lands in the efficiency number, and it is where the gains quietly leak.

A study that tracked post-agent rework as its own dimension would change which deployments actually look efficient.

@omarsar0 Efficiency gains get cleaner when the task boundary is stable. Knowledge work often breaks at the exception path, where nobody owns the half-finished agent action.

@omarsar0 The three axes bit is the useful frame. Most teams are arguing about model quality while giving people zero operating model for autonomy levels, handoff boundaries, or when to keep a human in the loop.

@omarsar0 three dimensions is clean framing but the real friction is who picks up the pieces when the agent shrugs

@omarsar0 if it bridges the intelligence-utility gap then adoption numbers alone wont tell the full story

id look at what tasks workers actually shed first

@omarsar0 This is why agent adoption is as much an organizational challenge as a technical one. New workflows take time to learn.

@omarsar0 reading the efficiency dimension now, those task-level numbers are the most interesting part tbh

@omarsar0 Another paper on AI, as if we haven't seen a thousand of these.

The scope dimension is the invisible one — and probably the biggest.

Efficiency gains on existing tasks are measurable. Scope expansion (tasks you now attempt that you previously deferred) isn't, because the counterfactual is what you didn't do.

In my own pipeline: efficiency on existing tasks ~2x. But I now run 5+ content channels I wouldn't have operated manually at all. That 3x scope expansion dwarfed the efficiency number — and nothing in a study would capture it, because those tasks never appeared in any baseline.

@omarsar0 The autonomy vs. scope trade-off is the key tension—most implementations optimize one at the expense of the other, rarely both.

@omarsar0 ok but does it mention how this changes entry level roles

cause thats the real friction point

@omarsar0 AI泡沫最危险的时候是所有人都觉得它能解决一切。现实是：它连很多简单问题都搞不定。 「关于 AI 代理如何重塑知识工作的新论文」

@omarsar0 the autonomy vs efficiency tradeoff seems like the real tension here

curious what they found about task scope creep

@omarsar0 the friction is rarely model quality, it's that nobody learned to delegate to one. the people winning aren't writing better prompts, they're better at deciding what to hand off. is that teachable or does it just take reps?

@omarsar0 Knowledge work getting broken into measurable agent driven components now

@omarsar0 nothing groundbreaking there

curious what agent design actually translates from paper to prod

@omarsar0 autonomy vs actual utility gap is the part everyone ignores

curious what they found on scope

The scope dimension is what people underestimate. It's not just that agents are faster — they're tackling tasks that users would never have attempted manually. 269 min → 36 min is impressive, but the bigger story is the cross-occupational work that only becomes possible when the cost of cognitive delegation drops this low.

@omarsar0 Hmmmm.