Key Takeaways
- When testing Anthropic's Claude Duet against products like Grok Bot and Muse, Claire Vo found that slow inference times ruined the synchronous user experience.
- Complex multi-step tasks, such as finding Golden State Warriors tickets across parallel threads, took over ten minutes to finish or ask for permissions.
- Users routinely prefer a model that is 10 percent less capable if it responds instantly, treating agents like colleagues who answer quickly rather than coworkers who disappear into meetings.
- Frontier reasoning models like Opus 55 provide high-quality output, but their slow response times push interactive products into asynchronous batch workflows.
- Builders must route synchronous interactive tasks to faster model tiers and keep heavy reasoning engines strictly for background processing.
Slow Models Turn Active Collaborators into Ignored Coworkers
Claire Vo tested Anthropic's Claude Duet on multi-step workflows, including buying Golden State Warriors tickets and coordinating parallel workstreams across artifacts and channels. Her main complaint was not about model accuracy. It was about speed.
“The second thing that I don't love about it, and this has been my experience with Claude generally, is it's slow, man,” Vo noted. “And I get so disengaged waiting for the models to finish.”
Vo compared the interaction to dealing with a coworker. In theory, an agent offers around-the-clock availability and prompt help. In practice, high latency breaks that dynamic completely.
“One of the benefits I think of an AI agent is that it is like a colleague, but a colleague who's up 24 hours a day and responds to immediately,” Vo said. “But if that colleague is slow, if that colleague goes to a meeting to get work done, like you kind of forget and don't want to go to that colleague over your more responsive ones, even if the more responsive ones do like 10% worse of a job.”
When an interface asks a user to collaborate live, every second of silence taxes working memory. If an agent makes a user wait two minutes before acknowledging an input or asking a clarifying question, the user leaves the tab. Once they leave, the collaborative session dies.
The Ten-Minute Loop Breaks the Interface
The friction becomes acute when an agent attempts parallel execution. In Claude Duet, complex requests spin off multiple threads to handle sub-tasks simultaneously. While that architecture sounds efficient on paper, slow inference turns it into a bottleneck.
Waiting ten minutes for an interactive tool turns an assistant into an interrupted batch job. Even top-tier reasoning engines cannot overcome that delay. Anthropic's Opus 55 model produces excellent technical output, but the execution time makes synchronous workflows painful.
“Even though I really love Opus 55 as a model and even though it does really good work, the performance is so slow and you really feel that in a personal agent and it degrades my desire to work synchronously with an agent,” Vo explained. Her verdict was clear: “The reality is like the latency and the inability to close loops meant that this particular agent product just really fell to the bottom of my use list.”
Speed Beats Raw Intelligence in Interactive Sessions
Founders and engineers often assume users will gladly trade waiting time for higher accuracy. Model leaderboards reward complex chain-of-thought benchmarks, reinforcing the belief that deeper reasoning always creates a better product. Real-world usage proves the opposite.
In an interactive workflow, speed enables iteration. If a fast model makes a small error, the user spots it in two seconds, adjusts the prompt, and moves forward. The conversation stays fluid. The human remains in flow.
When an agent uses a slow frontier model, that feedback loop shatters. The user cannot iterate rapidly because every adjustment carries a five-minute penalty. Instead of staying engaged, the user switches tasks, answers an email, or opens Twitter. By the time the agent completes its run, the user has lost the thread of what they were trying to accomplish.
What to Do With This
Audit your agent's response latency across every user touchpoint tomorrow morning. Split your architecture into two distinct pipelines: interactive synchronous loops powered by fast models that respond within five seconds, and asynchronous background jobs that notify the user via push notifications or email. If any live user workflow takes more than fifteen seconds to return output or request human permission, drop the task down to a faster model tier immediately.