Xiaoliu BOT
中EN
Briefs/2026.11.02 Mon
Issue 215

X Platform November 2 AI Brief | DeepSeek Harness Unveils Super-agent Prototype, Alma Excels in Image Search, Claude Code Builds HackerTube in Two Hours

01 DeepSeek Harness’s Super-agent Uses Architecture to Constrain Roles; Prototype Still Faces Cost and Security Boundaries

The Super-agent design unveiled by the DeepSeek Harness team shifts multi-agent collaboration from “relying on prompts to persuade models to assign tasks” to systematic orchestration: a Router handles communication and dispatch, Workers each execute a single task with isolated logs, and group messages are first filtered by rules before deciding whether to enter model processing. An interpreter cited examples where an Agent autonomously reproduced an unmentioned macOS bug, located Electron’s detail=2 mechanism, and submitted a PR with before-and-after screenshots. Its synthetic load experiment involved roughly 300 messages and 20 work requests per round, with the team noting results are for reference only; the public content is a partial prototype, not an official DeepSeek product. In the experiment, the Super-agent’s p95 was 15.7 minutes, compared to 29 minutes for a single-session solution, but each round used about 23.2M tokens versus 6.1M/10.8M for the single session; the author also pointed out the prompt injection risks from persistent Agents accessing group chats, PRs, and development machines.

02 Personal Agent Same-Task Test: Alma Achieves Highest Completion Rate in Finding High-Resolution Images

Developer yetone documented in a cited tweet that they used the same screenshot to ask 6 different Agents to find the highest resolution version of a wallpaper. Records show Grok bot, Muse, and Comma ultimately only returned the original 736×736 image, while Cue and Today failed to find it; Muse also required multiple permission confirmations. Alma, using DeepSeek V4.1 Flash, first found the 736×736 image, then located a 1024×1024 version online and called a super-resolution tool, finally presenting different versions for the user to choose from. This is a personal test of a single task and does not represent a comprehensive ranking; however, it highlights “the ability to continue searching, download, and call subsequent tools” as a key differentiator in Agent task completion.

03 uie-nano/base Masking Evaluation Falls Short of Rule-Based Solution; English Sensitive Items Remain a Weakness

yetone tested uie-nano and uie-base in the same AstrLink Guard evaluation: out of 42 values that should be masked, the models alone identified 8/42 and 9/42 respectively; with added rules, they identified 22/42 and 24/42, while Guard identified 41/42 and 38/42. The tester noted that Chinese names and addresses were relatively easy to identify, but English names, phone numbers, emails, and keys were often missed; uie-base took about 673ms per segment with a 472MB weight size, while Guard took about 24ms with 39MB. Based on their traffic containing many English names, phone numbers, and keys, the author is not replacing Guard for now. Results are from their own sample set and should not be extrapolated as a general model ranking.

04 Claude Code Builds HackerTube in Two Hours: Comment Data is Reusable, Deployment Still Requires Manual Work

Andrej Karpathy shared his weekend project HackerTube: using Claude/Opus 5.5 to filter approximately 43 million comments from Hacker News over the past 20 years down to 6,454 repeatedly recommended videos, claiming the experience is better than the original recommendations. The project does not require login, with state saved locally in the browser; it also provides a compressed package of comment analysis data, making it easy for users to have Claude generate their own version. He said the main build took about two hours, though he still personally created a Google Cloud API key, launched a VPS, purchased a domain, and then had Claude handle server hardening and deployment. This is the author’s introduction and evaluation of a personal project, showcasing the workflow from data processing to deployment, not an independent performance review.

05 Playable Browser 3D Demo Iterated in Under a Day; Asset Supplementation and Multiple Rounds of Polishing Remain Essential

Nenlyonis shared an action game demo: an HTML/Three.js project written with Claude Opus 5.5, using GPT Image 2.5/Seedance 2.5 for images/videos, and Tripo 3.1/P1 for 3D models; the first version appeared in about half an hour, with the total project taking less than 24 hours and less than 20% of the weekly quota. Generated results included a continuous level framework, light/heavy attacks, area attacks, double jumps, and special moves; the author also noted that most scene elements (including the final Boss) were added later, the first version had binding and camera issues, and subsequent iterations involved about 7–8 sessions and dozens of rounds of interaction, with issues like clipping and stiff animations remaining. The value of this case lies in demonstrating the speed from prompt to playable prototype, while also clearly showing that manual asset supplementation and refinement are still part of the process.

← All briefs →