Shadcn’s published task runs went from 69 UI errors before correction to zero afterward. The most useful example is small: a coding agent was asked to make a delete button stand out, so it added padding, raw pink styling, and a rounded shape.
Shadcn/lint pointed the agent to the button’s size and destructive variants instead. That is a more useful correction than telling an agent to “follow the design system” and leaving it to decide whether it did.
The Delete Button Got a Specific Fix
The lint error named the component settings the agent should use. Margin and width could still pass because they controlled where the button sat. The raw appearance changes did not.
That distinction matters when an agent is editing a UI. It can still place a button where it needs to go, while using the component’s intended settings for how the button looks. In the published example, the correction was specific enough for the agent to act on.
Zero Errors in These Runs Has Limits
Across over 150 task runs, Shadcn’s published table showed 69 errors before correction and zero after the correction rounds. For three Claude models, runs with diagnostics converged in one round. Shadcn also reported correction costs 10 to 48% lower than in the described control.
Those numbers are encouraging, but their scope is narrow. Per-model run counts were small, a Claude-family model judged the work, and the rules and prompts changed between measurements. The no-restyle rule also cannot read a class assembled at runtime.
So zero means zero errors after correction in the reported runs. It does not mean the rule catches every way an agent might restyle a component. The cost result has a similarly concrete next test: whether independent runs reproduce it.
One GPU Has to Balance Replies, Prompts, and Memory
The UI example is about giving an agent a correction it can use. Serving several conversations brings a different engineering trade-off: deciding how they share GPU time.
A model reads a prompt and then generates a reply one token at a time. With several conversations active, vLLM decides at each step which requests get GPU time. A waiting request can join when another reply finishes, and a long prompt can be split into chunks so it does not hold up conversations already receiving answers.
The token budget changes what users wait for. A smaller budget keeps the gaps between reply tokens shorter. A bigger budget gets long prompts through sooner. Neither setting removes the trade-off.
Long chats also take up cache memory. The vLLM walkthrough estimates about 57 kilobytes of KV cache per token for Qwen 2.5 7B instruct. That puts a 4K-token chat at roughly 235 megabytes and a chat at its 32K-token limit at about 1.9 gigabytes, before overhead and block rounding. The cache holds information from tokens the model has already read, so a longer chat occupies more of it.
vLLM’s launch post says older serving systems wasted 60 to 80% of KV-cache memory through fragmentation and over reservation. It reports under 4% waste with small cache blocks. Those blocks can be reused when a new turn repeats earlier chat history; if the cache fills, vLLM can pause a request and recompute it later.
Ongoing Assistants Have to Handle Calls and Purchases
The Muse examples show work that stretches beyond one reply. One speaker said Muse found 5,350 bucks a year in his subscriptions, of which he had canceled 1,285. Brad reported that Muse spent 98 minutes calling AT&T while contacting Verizon and T-Mobile to compare offers. A parent reported one daily notification combining schedules for four children in four sports across four messaging apps.
Claude Tag brings ongoing work into a team channel. It can use accessible chat history, documents, and tools to compile information or make updates. It works through its own accounts, logs changes and who requested them, and asks permission each time it responds on someone’s behalf. It is available for Team and Enterprise accounts.
Calls raise a question the speaker wants answered more clearly for Muse. Reuters reported that Meta tested human contractors handling some calls for employees, then rolled the experiment back after disclosure and privacy concerns. That report does not establish how Muse handles its calls. Amazon separately said Muse did not properly identify itself, accessed its store without agreement, and appeared to capture and store customer credentials. The speaker interprets that dispute as a fight over the customer relationship.
Completed purchases are another separate test. Walmart told Wired that checkout inside ChatGPT converted at about a third of its expected rate for products requiring a clickout. The speaker also says ChatGPT brought new customers at roughly twice the rate of search engines. Neither figure establishes what Muse’s retail partners will achieve.