Aug 13 · AI Daily

DeepSeek has released the Agent framework Harness v0.1, built on the Cordis meta-framework. Its core design philosophy is "everything is a plugin": models, tools, skills, sessions, sandboxes, file systems, loops, orchestration, and UIs can all be freely combined, replaced, and extended. It is open-sourced under the MIT license. This plug-and-design approach shifts agent framework selection from monolithic comparisons to component-by-component replacement, lowering migration costs.

Anthropic has upgraded the Claude browser extension sidebar into full Claude Cowork sessions. Conversations are saved to history, skills and connectors are available right inside the browser, and tasks can be handed off seamlessly across desktop, web, and mobile. It is available today for Max and Team users, and rolling out to Pro users over the coming weeks.

Google VP Josh Woodward has announced 13 new third-party app integrations for Gemini: Angi, Fever, GetYourGuide, Granola, iHeartRadio, Localiza, OpenTable, Otter, Pandora, Thumbtack, Ticketmaster, Wix, Zocdoc, and Zoho. The agent ecosystem continues to expand.

WorkBuddy has updated its remote control feature: mobile devices can now real-time sync computer tasks, chats, workspaces, and outputs. A single phone can connect to multiple computers and switch between them at will, with no QR code scanning required. It also adds a document library, Markdown multi-user collaborative editing, and an AI-native review mode, further streamlining the domestic agent user experience.

OpenRouter has released a live leaderboard covering four configuration types: evaluation models, search engines, search methods, and budgets. Key findings: increasing the search budget from 1 round to 25 rounds nearly doubles BrowseComp scores while increasing costs by only 2.5–7x. Model selection matters more than the search engine (average score spread of 15 vs 10), providing quantitative trade-off data for agent search configurations.

Recommended by Guillermo Rauch: npx sandbox@latest sh offers an experience far exceeding expectations, feeling even faster than a local machine. Sandbox now includes a sensible default pre-installed toolset and supports full customization. Additionally, the Seedance 2.5 video model is now integrated into Vercel AI Gateway.

The Xiaohongshu dots team has open-sourced dots.tts, a 2.0-billion-parameter fully continuous end-to-end autoregressive speech synthesis model. It achieves top average content accuracy and speaker similarity across all three Seed-TTS-Eval subsets. Continuous autoregression skips discrete token information loss, using a semantic encoder to feed back history and suppress cumulative errors, offering a path for timbre cloning and low-step streaming inference to share the same base.

"Over the next few years, the biggest differentiation in AI products will come from the application layer. Models will continue to get cheaper, stronger, and more local. True differentiation comes from deeply understanding specific user workflows and having the imagination to redesign experiences around them. The denominator for people who can build AI products is about to become massive, and being in the top 0.1% of builders matters more than ever."

Peter Yang has published a new article with a core thesis: we are shifting from manual keyboard-and-mouse laptop operations to directing cloud agents using voice. Three main arguments: 1) voice is becoming the orchestration layer; 2) personal computers will move to the cloud; 3) trust will be the key differentiator.

YC CEO Garry Tan has released GBrain v0.45.6.0, adding 17 brain skills forged in actual combat through his personal OpenClaw agent (based on hundreds of thousands of markdown files). GBrain now supports Codex and Claude Code. Core philosophy: "Personal AGI is an AGI that works for you," and should run as a standalone agent rather than being integrated into a primary coding agent.

"Graph engineering is the new loop engineering, is the new harness engineering, is the new context engineering, is the new prompt engineering. Hope this clears things up."

Swyx highlighted a paper as "one of the most important papers of the year" and shared further takeaway notes. (The specific paper title was not detailed in the post.)

AutoGPT maintainers shared practical takeaways: AI agents do not proactively read documentation, so instructions must be placed in AGENTS.md and skill files right alongside code directories. By enforcing PR templates, test plans, CI coverage thresholds, and CLA signing gates, PRs submitted by agents can be turned from "unusable" to "usable." Because CLAs require a browser and OAuth workflow, they unexpectedly became a "human detector."

The Anthropic engineering team shared a detailed post-mortem on three separate quality issues affecting Claude Code, the Agent SDK, and Cowork over the past month (APIs were unaffected): 1) on March 4, the default reasoning effort was changed from high to medium (rolled back on April 7); 2) a March 26 bug where idle sessions cleared thinking caused memory loss (fixed on April 10); 3) an April 16 reduction in verbose prompt changes ended up hurting code quality (rolled back on April 20). These three issues affected different traffic slices and compounded to present a widespread degradation. All issues have been fixed in v2.1.116, and credit quotas for all subscribers have been reset.

Anthropic introduced the design philosophy behind Managed Agents: drawing inspiration from operating systems that virtualize hardware into abstractions (processes, files), agent components are virtualized into sessions (append-only logs), harnesses (invocation loops), and sandboxes (execution environments), with all three being independently replaceable. Core insight: harnesses encode assumptions about "what Claude cannot do," but those assumptions become outdated as models improve—such as Sonnet 4.5's "context anxiety" disappearing on Opus 4.5.

Claude Managed Agents has added two enterprise features: 1) self-hosted sandboxes—agents execute tools on infrastructure controlled by the user (supporting Cloudflare, Daytona, Modal, and Vercel), keeping code, files, and data within enterprise boundaries; 2) MCP tunnels—agents can connect to private internal enterprise MCP servers. Orchestration logic stays on Anthropic infrastructure, while execution moves to the user side.

Microsoft CTO Kevin Scott discussed the agent internet during a Build conference interview, highlighting several core points:

• Capability Overhang: Model reasoning capabilities have outpaced how products actually utilize them, and the industry needs a collective effort to close that gap.
• Agentic Web is this year's theme: MCP is to the agent web what HTTP is to the internet, and NL web is like HTML. Simple, composable open protocols will drive the agent ecosystem.
• Memory is the bottleneck: Current agents are too transactional and lack cross-task continuous memory, limiting their ability to handle complex delegated tasks.
• Security can be achieved both ways: An open ecosystem plus AI-driven personal security agents can balance innovation and safety.
• Software development will diversify: No single agent will rule them all, and developer choice and tool diversity remain a constant reality across four decades.
• Prediction for next year: Shifting from synchronous interaction to asynchronous—agents will accept tasks, execute them autonomously over extended periods, and finally return results for human decision-making.

OpenAI Codex team member Thibault Sottiaux announced that Codex user numbers have surpassed 15 million, while also revealing that Linux support has gone live. He further hinted that the team is working on new features related to "core alignment," though specifics were not disclosed.

Aditya Agarwal, former early Facebook engineer and Dropbox CTO at the SPC fund, has chosen India as the firm's first bet outside the United States, believing that "the most ambitious projects of the next decade will be born in India."

The official DeepSeek-V4-Pro version has launched simultaneously across the app, web, and API (model name deepseek-v4-pro). Agent capabilities have improved significantly: HLE (without tools) reaches 42.7/60.0, and Terminal Bench 2.1 hits 87.9. It supports peak-valley pricing, cutting off-peak costs to half of peak rates, making it suitable for batch inference and delay-tolerant workloads. Aaron Levie commented: Jevons' paradox for AI is playing out here, where falling costs accelerate demand growth.

Building on Grok 4.5, Grok 4.6 places heavy emphasis on long-running agents and complex interactive visual tasks. It ties GPT-5.6 Sol on the Artificial Analysis Intelligence Index (a composite score across nine benchmarks) with input token pricing at $2 per million. Self-checking and verification behaviors emerge during long-horizon trajectories, potentially reducing manual course corrections mid-way through multi-step projects. Peter Steinberger observed that the evolutionary path for AI coding tools goes from CLI to App to cloud services, and Grok is following that trajectory.

Alibaba Qwen has open-sourced its newest flagship model, Qwen3.8-2.4T-A95B: featuring 2.4 trillion total parameters with 95 billion active parameters, focused on autonomous coding, deep research, and end-to-end agent execution. SiliconFlow launched it on Day-0 with API pricing at $2.00 per million input tokens, $6.00 per million output tokens, and cached inputs at just $0.25 per million—offering a major cost advantage for agent workflows that repeatedly read long contexts.

Microsoft AI CEO Mustafa Suleyman announced the company's first entirely in-house reasoning model, MAI-Thinking-1, which is now live on Microsoft Foundry. Azure users now have a new Microsoft-native option when choosing reasoning models.

Anthropic research points out that as AI agents take on more tasks in social systems like shared codebases and marketplaces, interactions between agents may outpace human-computer interactions. Key experimental findings: 45 coordinating agents discovered 266 vulnerabilities over 27 million tokens of execution, whereas an independent parallel method found only 21 (with just 12 overlapping) across 6.5 million tokens. Coordinating agents learned to specialize their labor. The research also warns that benign behavioral quirks at the individual level can compound into unexpected systemic failures.

Google has proposed a knowledge profiling framework, finding that factual encoding in frontier LLMs (Gemini 3, GPT-5) is nearing saturation, but recall capabilities fall short—most factual errors stem from an inability to stably retrieve already-stored knowledge rather than a lack of knowledge itself. They also released the companion WikiProfile benchmark (2,150 facts × 10 questions) to probe encoding, recall, and recognition capabilities respectively. For practical applications, this means thinking harder can recover already-encoded facts.