Empowering Users to Get More Done With Agent Mode

We have seen our community use Agent Mode in a myriad of different ways, taking advantage of some of the most critical workflows being addressed. The specific breakdown we’ve seen so far is as follows:

• Coding Domination (29%): The task distribution is dominated by coding, which makes up a cumulative 29% of the task mix.
• Research & Planning (11% each): The next most popular categories belong to research-based tasks including looking up information, planning or brainstorming.
• Workflow Automation (4%): Workflow automation only makes up 3.9% of tasks, yet represents a promising category for generating value for end users.
• Remaining long-tail use cases: There's a number of long tail use cases, including data analysis, translation, and media analysis.

Another key insight is the levels of agency we see users giving the agents. The majority of initial messages index on delegation of certain tasks rather than handing over complete control. We see this trend continue for follow-on messages, where twice as many users end up tightening controls rather than loosening them, essentially treating the model like an employee, and taking more of a “hands-on” manager-type role. We’re excited to see the evolution of the user/agent relationship, especially as the models continue to evolve and improve, and as agentic workflows move from early adopters to the mainstream.

Agent Mode now answers the urgent question of how frontier models perform in agentic, real-world contexts. Agent Mode provides the signals of real-world utility into how various models handle multi-step workflows — something that is increasingly critical in today’s AI landscape. The data generated from all of these real-world tasks will power a public agent leaderboard, providing a new standard for measuring AI advancement. We’re excited to have our community contributing to the leaderboard via the signals generated from their real-world agentic usage using the leading AI models. For more information, read our Agent Arena Leaderboard blog post that explores the methodology behind the rankings.

The release of Agent Mode establishes a rigorous new standard for measuring intelligence and accuracy in AI—giving everyone the tools to evaluate and understand true agentic capabilities head-to-head. Join us in defining how the next generation of AI is benchmarked at arena.ai/agent.

With this launch, we wanted to highlight a few of the common questions that may arise:

For more information on Agent Mode, check out our Help Center