01 Leaked Claude Code shows Anthropic experimenting with persistent agents and a hidden ‘Undercover’ mode
A publicly posted package from Anthropic’s Claude Code — apparently released with the 2.1.88 update — included a source map that exposed portions of the TypeScript codebase. Reported excerpts and coverage say the leak contains hundreds of thousands of lines of code and reveals features that aren’t prominent in the product documentation.
Among the exposed artifacts are references to a persistent, always-on agent and an experimental mode labeled “Undercover,” along with UI concepts such as a Tamagotchi-style virtual companion named Buddy. The materials in the package give a rare look at how Anthropic is structuring background agents, agent lifecycles, and agent-facing UI affordances.
The leak’s contents are notable because they show design choices — always-on processes, persona-like companions, and stealth modes — that could shape how developers build agent-based workflows atop Claude Code. Public reporting is limited to the files that surfaced; Anthropic’s official product descriptions and road map remain the primary source for claims about shipping behavior and user controls.
- A Claude Code package tied to the 2.1.88 update included a source map exposing TypeScript code and internal feature references.
- Leaked files reference a persistent always-on agent, an “Undercover” stealth mode, and a Tamagotchi-style virtual companion called Buddy.
- The artifacts provide insight into Anthropic’s agent design but do not by themselves confirm which features will ship or how controls will work in production.
02 Alexa+ gains conversational ordering for Grubhub and Uber Eats
Amazon is expanding Alexa+ to let users place and modify orders from Grubhub and Uber Eats through a conversational flow. The company frames the experience as similar to ordering at a restaurant: users can build an order interactively and adjust items without navigating rigid menus.
In practice, Alexa+ will guide users through selections, let them change items (for example, add or remove sides), and confirm details in natural-language turns rather than forcing step-by-step prompts. Amazon says the goal is to avoid the awkward, tedious exchanges that have plagued voice ordering in the past.
The integration represents another push to make Alexa a higher‑value assistant for everyday tasks by connecting voice-driven dialog to third-party commerce partners. Rollout details and availability will depend on region and partner arrangements, and Amazon positions this as an enhanced capability for Alexa+ subscribers.
- Alexa+ now supports conversational ordering with Grubhub and Uber Eats, intended to let users build and change orders like a waiter-driven dialog.
- Amazon emphasizes natural-language turns to avoid the awkward exchanges common in older voice-ordering flows.
- Availability will vary by region and depends on partner integrations; the feature is tied to Alexa+.
03 OpenAI’s Gradient Labs outfits banks with AI account managers using GPT-4.1 and GPT-5.4 mini/nano
OpenAI announced Gradient Labs, a project that deploys AI agents to automate banking support workflows. The service uses a stack including GPT-4.1 and GPT-5.4 mini and nano models to deliver low-latency responses and higher reliability for routine account-management tasks.
Gradient Labs’ description emphasizes agent-driven automation at scale: agents handle support workflows and aim to reduce friction in common banking interactions. The blog notes the combination of model sizes is chosen to balance responsiveness and cost for production deployments.
For financial services, the practical implication is a push toward automated, conversational support tailored to account tasks. Details about partner banks, rollout timing, and customer opt-in were not enumerated in the summary material provided.
- Gradient Labs uses GPT-4.1 and GPT-5.4 mini/nano to power AI agents that automate banking support workflows.
- OpenAI highlights low latency and reliability as design goals for these agent deployments.
- Public materials do not list specific banking partners or a detailed rollout schedule.
What moved around the edges
Anthropic source release accidentally exposed TypeScript bundle and a 512,000‑line artifact
A package distributed with Claude Code 2.1.88 included a source map that made its TypeScript sources readable; reporters say the exposed files amount to more than 512,000 lines of code, revealing internal UI and agent components.
The Verge AIYupp shutters less than a year after raising $33 million
Crowdsourced model‑feedback startup Yupp announced it is closing operations after raising about $33M from investors including Chris Dixon, according to TechCrunch.
TechCrunch AIGoogle publishes a March 2026 AI product recap
Google’s AI blog posted a March 2026 roundup detailing the company’s recent AI updates and product changes across its services in a single recap post.
Google AI BlogArs Technica interrogates Anthropic’s method for measuring AI ‘theoretical capabilities’
Ars Technica reviews a 2023 Anthropic study that made assumptions about future LLM‑powered software, questioning how the company quantified AI’s hypothetical job‑market impact.
Ars Technica AIRoadie: an $86 open‑source KVM built for agent‑driven phone and tablet control
The Roadie GitHub project documents an open‑source hardware KVM that uses HDMI capture, USB HID out and a browser‑accessible HTTP/WebSocket API, supports multi‑touch for phones/tablets, and lists an estimated parts cost of roughly $86 with no software required on the target device.
Hacker News AI