Cerebras says its rehearsed AI agent booked dinner in 22 seconds

Cerebras reported a 22-second median across two successful runs of a dinner-booking AI agent using Qwen 3.8 27B on its inference hardware with Pi as the agent harness, which the company says is 19 times faster than Meta Muse, Claude Cowork and Grok Bot, whose recorded runs took 4 minutes 36 seconds, 6 minutes 25 seconds and 7 minutes 40 seconds. Cerebras said the result followed a rehearsed setup that taught the agent the booking route in advance and cut tool calls by more than 80%, and the company did not report comparative failure rates, costs or repeated trials across other websites and tasks. The single-task, company-run test does not establish a general advantage.

Cerebras says its rehearsed AI agent booked dinner in 22 seconds Cerebras reports a 22-second median across two successful runs using Qwen 3.8 27B, its inference hardware, parallel checks and a saved website-navigation skill. By RuntimeWire Staff https://runtimewire.com/author/runtimewire-staff · Published Primary source: Cerebras Newsroom https://www.cerebras.ai/blog/the-rise-of-slow-personal-assistants Why it matters Cerebras is making the infrastructure case in a consumer workflow: agent speed depends on inference latency and on how much unnecessary discovery and serial tool work the software performs. Its 22-second result is suggestive, but the small, company-run test does not establish a general advantage. On September 24th, Cerebras https://www.cerebras.ai/?ref=runtimewire , the AI-computing company co-founded by Andrew Feldman https://investors.cerebras.ai/board-member/andrew-feldman?ref=runtimewire , published a test of personal AI assistants https://www.cerebras.ai/blog/the-rise-of-slow-personal-assistants?ref=runtimewire , arguing that they can spend more time waiting and figuring out websites than doing the task. Cerebras' dinner-booking agent finished in 22 seconds, the company says, after its team gave the agent a faster model, parallelized restaurant checks and taught it the booking route in advance. That is a useful demonstration of the case Cerebras wants to make for fast inference: response time is part of whether an AI agent feels useful, especially when a task requires repeated model decisions and browser actions. Feldman's background helps explain Cerebras' focus. Before Cerebras, he led dense-microserver maker SeaMicro, which AMD acquired, and later served as an AMD corporate vice president. He holds a B.A. and an M.B.A. from Stanford. Cerebras builds AI compute infrastructure. Its September test takes that infrastructure pitch out of the data center and into a familiar consumer chore: reserve a table. Cerebras says its test assistant used Qwen 3.8 27B running on Cerebras hardware, with Pi https://pi.dev/?ref=runtimewire as the agent harness. Pi managed the conversation and tool calls; Cerebras added tools and a reusable set of instructions for the booking process. The faster run had a head start Cerebras compared the same dinner request https://www.cerebras.ai/blog/the-rise-of-slow-personal-assistants?ref=runtimewire across Meta Muse, Claude Cowork and Grok Bot. Its recorded successful runs took 4 minutes 36 seconds, 6 minutes 25 seconds and 7 minutes 40 seconds, respectively. Each made a reservation at one of three restaurants: A Mano, II Borgo or Doppio Zero. Cerebras says the 22-second result is the median of two successful runs of its own assistant. Cerebras labels the result 19 times faster than the other assistants. The comparison is a single task, not a general ranking: the assistants took different paths, made different numbers of tool calls and were not shown with fully specified model, hardware or software configurations. Cerebras also does not report comparative failure rates, costs or repeated trials across other websites and tasks. Its figures describe the runs it recorded, rather than an independently replicated benchmark. The optimization also started before the clock. Cerebras' team had already turned its knowledge of the booking website into a reusable skill, so the assistant did not have to discover each step from scratch during the timed run. Cerebras says that reduced tool calls by more than 80%. Availability and booking details still had to be checked live; the saved instructions covered the route through the site, not the changing information on it. The prepared workflow limits what the 19x figure can establish. In a recorded Grok Bot trace, Cerebras says one browser run spent 2 minutes 18 seconds checking three restaurants sequentially. Its optimized browser and API portion took 6.8 seconds, against 4 minutes 31 seconds in the earlier Grok trace. Cerebras explicitly says the comparison bundles changes to the model, harness and execution path. It does not isolate the impact of the saved skill or show that the same improvement would hold with the other assistants' full configurations. A compute pitch with a workflow attached Independent checks can run at the same time, while the booking decision still waits for their results. Faster inference can shorten the pauses between a model receiving a page, deciding what to do and calling the next tool. It cannot make a slow website load instantly, remove a required phone check or guarantee that an agent will recover from an error. That combination of compute and orchestration is central to Cerebras' argument. In August, Cerebras introduced its CS-4 system, and it has positioned its infrastructure around high-speed inference. RuntimeWire recently reported on Cerebras' CS-4 rack plans and its next system target in our August coverage https://runtimewire.com/article/cerebras-cs-4-nexus-rack-cs-5-token-speed . The dinner test applies the same speed proposition to a consumer workflow, where the measure people care about is whether delegation saves time overall. The shift is timely for Cerebras as it expands its public story beyond chip performance. In February, Cerebras announced a Series H financing https://cerebras.gcs-web.com/news-releases/news-release-details/cerebras-systems-raises-1-billion-series-h?ref=runtimewire of $1 billion at an approximately $23 billion post-money valuation, with Tiger Global leading and Benchmark, Fidelity, AMD, Coatue and other investors participating. Cerebras began trading on Nasdaq in May, according to its company history https://www.cerebras.ai/company?ref=runtimewire . A demonstration of faster inference in an agent workflow gives the infrastructure pitch a concrete outcome to point to, though this small test does not establish a broad performance advantage. For builders, the experiment is a reminder that an agent's elapsed time is shaped by more than the model's response speed. A workflow that repeats discovery, waits for serial checks and calls the model at every step can remain slow even when the model is capable. Reusing a tested procedure and running independent work in parallel can cut avoidable waiting; a faster model can shorten the remaining decision loops. Cerebras has shown one way to combine those changes, on one reservation task, with a prepared workflow and two successful runs. Whether the result carries across more complex errands, unfamiliar sites and failed attempts is the test that would turn the demonstration into evidence about personal assistants generally.