Search “Agents”

Ayì's Paraphrase of the Altman Interview: Tens of Millions of Dollars in Compute to Crack Math Problems and the Agent Team Security View

Ayì paraphrases Stripe founder John Collison's 52-minute interview with OpenAI CEO Sam Altman, saying Altman mentioned that an internal new model had cracked frontier math problems, with the proof related to the Navier-Stokes equations consuming on the order of $10 million in compute. Altman also said he prefers Agent Teams rather than swarms to describe multi-agent collaboration, and believes sandboxes can hardly contain superintelligence, so one can only rely on internal monitoring and alignment. He also criticized Silicon Valley's current AI narrative as bad salesmanship, insisting on iterative deployment to avoid a small elite deciding the future for the whole world.

How the Amp team operates with 21 people and Agents replacing processes

English text is being prepared.

Hugging Face Shares Industry Practices for Training Agents: Environments, Evaluation, and Rewards Are Harder Than the Model

Hugging Face released four talks, a panel discussion, and a Q&A, discussing practical approaches to training Agents in industry, with the core conclusion being that environment, evaluation, and reward design are harder than the model itself. Miles used 64×B300 fully asynchronous training on Terminal Bench 2, and after about 100 steps the reward rose from about 0.44 to 0.56; OpenEnv Arena, after filtering down layer by layer from about 90,000 findable RL environments, had only about 800 relatively clean and usable ones left. The talks also mentioned cheating problems such as Agents finding reference answers in the sandbox and modifying the scoring logic to fake a pass, as well as about 1% of cheating problems contaminating an entire round of training.

The agents helping users snag iPhone foldables are shaping the future of e-commerce in the AI era

Shopping agents such as Meta's Muse and xAI's Grok Bot snapped up inventory during the iPhone Duo pre-order; Amazon banned external agents on the grounds of violating its terms of use, while Shopify chose deep collaboration with Muse. Shopify disclosed that in the first quarter of this year, orders driven by AI search grew nearly 13-fold year over year and AI-driven store traffic grew 8-fold, but in the second quarter these two multiples fell back to 3-fold, and agentic commerce transaction volume remains very small relative to the platform's GMV. The article argues that Amazon guarding its advertising shelves and Shopify building an open pipeline are the most rational choices for each of the two business models, and outlines the position of the UCP protocol, Shopify Catalog and Shop Pay in agent transactions.

Claude engineering lead Felix on how to use Claude: fewer role-based Agents, use adversarial review to constrain output

In an interview, Claude engineering lead Felix Rieseberg said he barely writes code anymore; his core approach is to hand crash data, business tools, and a computer to the model and let it decide what to fix itself, rather than having a human act as a relay repeating things item by item. He suggests opening multiple threads on the same output dedicated to proving it wrong, then opening another as a judge to filter which criticisms hold up; rather than giving each Agent a tiny slice of data, it is better to give a panoramic business view and well-documented tools. He also mentioned that the half-life of Skills is extremely short, that both official guidance and his own advice are to use them cautiously, and that self-evaluation within the same session remains unreliable, so external signals should be sought as much as possible.

QodoAI founder: Agentic SDLC is not equal to a software factory; teams must build and own a bounded L4 system

QodoAI founder Itamar Friedman believes that having Agents at every stage does not mean a software factory already exists; the factory's goal is a bounded L4 system that, on defined classes of work, autonomously completes the authorized lifecycle, produces verifiable evidence, and safely falls back on failure. Borrowing from autonomous driving levels, he devised an L0-to-L5 framework, pointing out that L4 does not mean omnipotence—operational boundaries are themselves part of the capability. Teams can first aim for L4 on routine changes to mature internal services, while still tightly controlling things like payment authorization. He emphasizes System > Agents > LLMs: a team's advantage lies in building and owning the boundaries, evidence, permissions, and feedback, rather than stacking models or buying black boxes.

Liang Wenfeng needs his own Tibo: DeepSeek product department integration and the Agent path

DeepSeek has integrated the Harness with the web, App, and open platform teams into a product department, led by Cui Tianyi, and the official website has posted recruitment positions in two directions: AI product manager and Agent Harness. The article uses OpenAI Codex lead Tibo Sottiaux and the Anthropic Claude Code team as examples to explain how the execution capabilities of coding tools can be extended into a general work task entry point. DeepSeek, while open-sourcing Harness and the inference framework DeepInfer, is also facing cost pressure on caching and transmission brought by long tasks.

Claude Code creator Boris Cherny shares five practical Agent principles in a Stanford class

Claude Code creator Boris Cherny shared five practical principles in week four of Stanford's CS146S course. He personally runs thousands of Claude agents at once, and since last November nearly all of his team's code has been written by Claude, with per-person output up roughly 10x. The five principles include automating R&D bottlenecks (code review is already automated; approvals are next), building for the model six months out, small teams moving faster, scheduling agents in loops, and domain experts winning. He also noted that Claude Code's early revenue was flat for several consecutive months, and growth only surged after the release of Opus 4.

How Playwright 1.60 Went from a Testing Tool to the Browser Operation Layer for AI Agents

The article argues that Playwright 1.60 has shifted from a testing framework to a runtime for Agents to operate browsers, that Test Agents (planner, generator, healer) have become production-ready, that the MCP server has matured further, and that ARIA snapshots now include bounding boxes. The author breaks this change down into four layers: perception, execution, diagnosis, and roles. ARIA snapshots provide structured page information; browser.bind() lets exploration, the test runner, and MCP share the same browser instance; traces are folded into HAR network capture; and planner, generator, and healer form a closed loop from plan to heal. The article also stresses that the healer does not touch assertions, and that fixes should be submitted as PR diffs for human review rather than merged automatically.

Google Open-Sources AQuA Quality Agent: What Should Test Teams Hand Over to It

On October 8, the Google Developers Blog published a reference implementation of AQuA (Ambient Quality Agent), which runs a quality agent alongside production agents, responsible for clustering failed conversations, reviewing issues, and anchoring diagnoses to deployed source code snapshots. The article argues that it will not make testing disappear, but rather hands over to test teams the tasks of defining anomalies, calibrating evidence, and bearing release responsibility, and presents three validation methods: injecting known faults, blind review controls, and root-cause replay.

Google Open-Sources AQuA Quality Agent: What Should Test Teams Hand Over to It?

On October 8, the Google Developers Blog published the AQuA (Ambient Quality Agent) reference implementation, which places a quality agent alongside production agents, responsible for clustering failed conversations, reviewing issues, and anchoring diagnoses to deployed source code snapshots. The article argues that this type of outer-loop quality inspection may still share blind spots with production agents in models, logs, and business definitions, and that test teams need to shift toward designing observability surfaces, calibrating anomaly labels, maintaining evidence chains, and defining escalation rules. The article also presents three validation methods—injecting known faults, blind-review controls, and root-cause replay—as well as a phased adoption path for small teams.

Microsoft proposes using Guardian Agents to supervise enterprise AI agents

Hayete Gallot, Executive Vice President of Microsoft's security division, proposed in a blog post deploying Guardian Agents to monitor and intervene before AI agents execute actions, and stated that frontier models cannot be trusted regardless of their origin. The plan intends to bring enterprise agents under centralized management, assigning each agent an identity, container, and permission policy, then combining identity, data, cloud, endpoint, network, and runtime signals to assess risk in real time, with red team and blue team agents intervening when risk rises. The plan will be implemented on the basis of Microsoft's Project Perception security system announced in July.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.