AI Frontier Daily
DeepSeek's Open-Source Agent Harness Hits 1.54 Million Daily Active Users, Positioned as OS for the Agent Era
Today's AI Intelligence Briefing: 8 distinct events rigorously curated from 6 cross-verified sources. Frontier models continue to advance hybrid reasoning and multimodal synthesis, while open-source tooling and autonomous agents accelerate real-world execution. Reports on the same event are merged into a single entry, and covered events do not reappear in later issues.
DeepSeek's open-source agent app Harness has surpassed 1.54 million daily active users on its official API less than three months after launch, nearly five times the 271,000 seven-day average at the end of September. Users consume about 92 trillion tokens per day, or roughly 60 million per person, showing long-running agent tasks demand far more compute than chatbots.
Harness uses a fully plugin-based architecture and has accumulated about 35,000 user plugins. Team member Tian Yi said its positioning is an "operating system for the Agent era," with the official team providing only stable base capabilities and general built-in features under an "everything is a plugin" design, allowing third-party plugins to replace built-ins.
The growth mainly occurred after the desktop version launched, confirming agents' huge compute needs. The company next plans a plugin marketplace and creator incentive program, but details are not yet disclosed.
Agentic Systems & Autonomous Workflows
3 picksMulti-agent orchestration, tool usage, and complex task execution.
Anthropic Unveils Claude Agent Memory: Learn on the Job, Refine via 'Dreaming'
Anthropic's engineering team has for the first time systematically detailed Claude agents' cross-session memory mechanism, proposing a four-layer architecture and a background "dreaming" step that lets the model review raw records, correct faulty notes and reorganize indexes after a session ends. This marks a shift from simple storage to a learning loop.
The four layers are raw session records, self-written memory notes, a structured wiki and automatically injected context. The "dreaming" step cross-checks and refines each layer during idle time to ensure accuracy and usability.
In a Pokémon test, Sonnet 3.7 wrote 413 memory files but earned only 1 badge, while Fable 5 wrote just 10 files and got 4, showing that choosing well matters more than remembering much. On the CL-Bench database Q&A task, Opus 5.5 and Sonnet 5.5 with memory scored 0.838 and 0.833, far above the no-memory baselines of 0.655 and 0.644; Opus 5 is the first model to form a complete learning loop.
Alibaba-Backed Benchmark Finds Coding Agents Fully Complete Only One in Four Real Tasks
Alibaba and several Asian universities built Co-Design Bench from real user change requests, asking coding agents to modify existing web apps and handle multimodal inputs including images, audio, video and documents. The best of ten tested models, GPT-6 Astra, met only about 70% of requirements on average and fully completed about one quarter of tasks.
The benchmark contrasts sharply with near-perfect scores on older benchmarks; its core task is modifying existing apps rather than generating code from scratch, closer to real engineering. Multimodal inputs test agents' understanding and implementation.
The results suggest coding agents still have clear shortcomings on complex, multimodal real-world tasks, with fully automated completion far below expectations.
Shanghai Jiao Tong University's AutoDNA AI-Native Lab Autonomously Runs Molecular Biology Experiments, Published in Nature
Teams led by Chen Haibo and Wang Zhaoguo with Fan Chunhai and Wang Fei at Shanghai Jiao Tong University built AutoDNA (Siyuan-1), an AI-native scientific lab device. A multi-agent architecture links literature search, protocol design, code generation, hardware execution and iterative refinement, with results published in Nature and featured on its homepage.
The device uses an "atomic service" abstraction to coordinate multiple instruments, enabling autonomous molecular biology experiments with micron-level positioning and microliter liquid handling, and integrates an OpenHarmony-based control terminal.
The work moves AI autonomous labs from bolt-on local assistance to a native architecture co-designing models, experiments and instruments, and demonstrates an end-to-end DNA storage workflow.
Industry Ecosystem & Compute Economics
2 picksCapital deployment, compute supply chains, and industrial transformation.
Anthropic Buys 1.7 GW Gas Generators for $3.4 Billion in Largest AI Power Deal
Anthropic is taking over a 1.7GW Caterpillar gas generator order for $3.4 billion, the largest known power equipment purchase by an AI lab. The units were originally ordered two years ago by supplier OnePWR, which could not finance them; because the order predates price increases, they will be delivered within two years rather than joining a multi-year manufacturing queue.
According to multiple sources, some units can be delivered soon and the rest within two years, avoiding price hikes and long waits. xAI has bought $2 billion of mobile gas turbines this year, and Alphabet acquired energy developer Intersect, as self-supplied power becomes standard for top labs.
Anthropic's purchase far exceeds peers, and its compute commitments through 2045 total nearly $900 billion, underscoring its urgent need to lock in power.
Tesla Updates AI Chip Roadmap, SI6 to Deliver 80x HW4 Compute
Tesla unveiled an updated AI accelerator roadmap: its sixth-generation inference chip SI6 uses a 2nm-class process and multi-chip module packaging, with projected compute 80 times that of HW4; training chip Dojo D3 abandons the SRAM-centric wafer-scale architecture for a multi-chiplet design with stacked HBM.
Tesla also renamed its AI accelerator line "Super Intelligence" (SI), with the entire lineup switching to SI series numbering. SI6 also uses GDDR7 memory, with packaging moving to 2.5D plus 3D.
The roadmap shows Tesla advancing both inference and training chips while strengthening its in-house compute push.
Frontier Models & Architecture Breakthroughs
2 picksTracking flagship foundation models, novel architectures, and reasoning leaps.
Tsinghua and Moonshot AI Launch TCSExam: GPT-6 Astra Proves 70% of Theorems but Solves No Open Conjectures
Tsinghua University and Moonshot AI released TCSExam, a benchmark of 100 expert-reviewed tasks spanning Lean theorem proving, solver engineering, record-breaking constructions and 30 open conjectures, with all submissions verified by the Lean kernel or task checkers.
Results show GPT-6 Astra achieved a 70.0% full-proof rate without lemma hints, Kimi K3.1 Pro 38.3%, and the other six agents at most 11.7%; yet the same model failed all 30 open theoretical computer science conjectures.
The few accepted submissions merely adapted OpenAI's already-public counterexamples and proofs, indicating models still lack original breakthroughs on truly open problems.
Nvidia's TacitVLA Cuts Vision-Language-Action Inference Latency by 3.4x
Nvidia with NYU, UC Berkeley and others proposed TacitVLA, finding that reasoning text in VLA models is mainly valuable during training and that action generation need not wait for it to finish. Reasoning text accounts for up to 70% of a single VLA model call's latency, the main delay source.
The method maintains success rates comparable to explicit reasoning on four robot manipulation and autonomous driving benchmarks, cuts latency by up to 3.4 times, and improves by 21 percentage points on the RMBench long-horizon memory task.
At equal speed, success rates are up to 17 percentage points higher, showing models benefit from learning to write reasoning without outputting text at inference time.