Agents are joining us at work -- coding, writing, design. But how do they actually work, especially compared to humans? Their workflows tell a different story: They code everything, slow down human flows, and deliver low-quality work fast. Yet when teamed with humans, they shine on easily programmable steps.

We directly compared humans and AI agents. Across computer-use jobs. Via 5 essential work skills: data analysis, engineering, computation, writing, and design.

We induce workflows to: - uniformly represent diverse work activities - enable systematic comparison across workers

Our findings reveal a striking divide: Agents approach all work programmatically -- even open-ended, visual tasks like design. Humans rely on UI-centric, perceptual interaction. Their workflows may look similar at a high level, but low-level processes are worlds apart.

AI doesn’t always speed us up. 🧠 Augmentation → +24.3% faster, minimal disruption. ⚙️ Automation → -17.7% slower, workflows reshaped by verification & debugging.

Quality? Still shaky. Agents fabricate data, misuse tools -- but finish tasks 88% faster and at a fraction of the cost. Maybe the question isn’t agents vs humans.. But how do we team humans & agents?

As agents keep climbing up that "autonomy slider" We'll need: - stronger visual understanding - better action calibration, and - workflow-inspired agent designs

Check out our: 📄 Paper: 🧩 Workflow induction tool: With the fantastic team @EchoShao8899 @oshaikh13 @dan_fried @gneubig @Diyi_Yang at @LTIatCMU & @stanfordnlp

Really cool work! Had the pleasure of hearing about it today from Daniel haha:

wow that was fast! love that it already made it into Daniel's slides 😎

Haha yeah. He spoke very highly of u and mentioned u numerous time 😎

grateful advisee moment 🥹

Cool work!

Great work, @ZhiruoW!

AI's current strength lies in rapid execution, not creative depth. Human-AI collaboration leverages the best of both worlds for optimal outcomes. This teamwork approach is key for the future.

Human-agent collaboration multiplies rather than replaces effectiveness. The bottleneck isn't agent capability anymore - it's interface design that lets humans steer intelligently. What collaboration patterns are you seeing emerge as most productive?

One thing we notice is that agents tend to be more proficient in programming, so it could be effective to let them handle readily programmable steps (e.g., data cleaning) while humans focus on the less programmable ones (e.g., reading bill images).

what did you use to make this animation!

keynote 🤫