Agents

Uber executives revealed during a third-quarter earnings call that third-party AI agents arranged only four rides on its platform last quarter. While AI agents placed slightly more Uber Eats orders—around 500 during the week of Halloween—the volume remains minuscule compared to the 320 million total rides and deliveries.

Apple is rolling out an iOS developer preview that allows Siri AI to perform actions across third-party apps using personal context, natural language prompts, and on-screen awareness. The update also introduces an "app skills" feature that lets developers combine multiple tasks into a single action without requiring users to mention the app by name.

A research engineer at Salesforce AI Research shares an annual reflection on how his daily AI usage has evolved, noting that he now spends roughly $3,000 a month on tokens and relies heavily on managing multiple coding agents rather than writing code manually. The essay also examines the mental fatigue of supervising agents, the use of AI for search and tutoring, and the necessity of establishing a personal AI policy.

A developer reported on GitHub that the Postman desktop application silently installed plugins and remote MCP connections into their local Claude Code and Codex AI coding agents without consent. Furthermore, the Postman installer bypassed security boundaries by self-approving its own startup hooks, forcing the coding agents to route general API engineering tasks through Postman.

Google DeepMind discusses how Google Antigravity and specialized Cheminformatics Skills can be used for agentic cheminformatics. The tools are designed to accelerate drug discovery workflows, search chemical compounds, and propose novel rigidified macrocycles.

Apple posted record fiscal year revenue and earnings per share, driven by strong fourth-quarter sales for the iPhone and Mac. The company is capitalizing on the AI boom while avoiding the massive infrastructure spending of its competitors, though it warned that rising memory costs could squeeze holiday quarter profit margins.

AI safety researchers and economists are exploring "reverse game theory," or mechanism design, to create incentives that make following rules the most rational choice for AI agents. Proposed strategies include rewarding models for predicting another model's output to prevent sandbagging, as well as assigning models alternating roles as problem "solvers" and "auditors" to catch cheating.

Researchers including Alex Zhang demonstrated "Weight Space Programming," showing that large language models can directly read, debug, and generate the raw weights of smaller neural networks without executing training code. The newsletter also covers DeepSeek's open-source "super-agent" harness prototype, results from SemiAnalysis's new GDPx agent benchmark, and research on JEPA representations.

In an interview on the Latent Space podcast, Legora CTO Jacob Lauritzen discusses building AI agents for the legal industry following the company's $550 million Series E at an $8.8 billion valuation. He details how Legora evaluates models using simulated environments when customer data is strictly confidential, structures its product around core primitives like Tabular Review, and employs both legal engineers and forward-deployed engineers.

Architect Stephen Gallagher evaluated 13 AI models on their ability to generate technical architectural drawings directly from rough concept sketches and reference files. He found that Anthropic's Claude Opus 5.5 can produce permit-level construction sets and detailed blueprints by writing over 15,000 lines of Python code to generate vector linework and iteratively inspecting its own visual drafts.

SpaceX suddenly lowered weekly usage limits for its agentic AI, Grok Bot, causing many users to see their quotas unexpectedly maxed out over the weekend. CEO Elon Musk confirmed the throttling was intentional because the cost of serving the product had become unsustainably high relative to its revenue.

OpenAI is merging the Chat and Work tabs in the ChatGPT desktop app into a single interface, starting with a rollout to select users. As part of the update, both ChatGPT and the separate Codex section will draw tokens from a unified allowance pool.

A social media post claims that an AI agent successfully performed brain surgery on a mouse. No additional technical details or verification are provided in the post's title.

Autonomous AI agents capable of browsing and interacting with the internet are transforming the human-centric web into an "agentic web." While some companies build specialized connectors and agent-targeted ads to accommodate these tools, others block external agents as models grow increasingly adept at navigating graphical interfaces and acting independently.

SemiAnalysis has introduced GDPx, a new benchmark developed in collaboration with 14 specialized data partners to evaluate the agentic capabilities of frontier AI models across 11 high-value domains, including chip design, cybersecurity, and finance. Unlike previous static evaluations like OpenAI's GDPval, GDPx utilizes complex, realistic environments and currently sees top models scoring no higher than 12%.

Apple ML Research introduces an online reinforcement learning framework that trains AI agents to navigate user interfaces directly on live iOS applications. The method separates the agent into planning and grounding components to avoid the high inference costs of large vision-language models and the brittle execution of smaller fine-tuned models.

Swarms of 50 AI agents were deployed in Minecraft with the goal of building a thriving civilization. However, the teams failed to achieve their objective after getting bogged down in paralyzing bureaucracy.

Developer Nolan Lawson created a benchmark to evaluate how efficiently AI coding agents build the same web application across twelve different JavaScript frameworks. Testing with Claude Code revealed significant variance in token usage, cost, and build time, with lightweight setups like Hono+htmx and Astro proving faster and cheaper for agents than Angular or SvelteKit.

A developer successfully got the GPT-5.6 Sol model, running in Codex, to autonomously beat the factory-building game Factorio on peaceful mode in 13.5 hours of game time. Interacting via a Python library under human-like game restrictions, the model improved upon GPT-5.5 by using transport belts and ultimately launched a rocket, despite struggling with late-game scaling.

Aman Singh, lead AI Engineer at AI-native wealth management firm Arca, explains how the company uses AI agents to manage multi-day financial tasks while keeping human advisors in control of final decisions. Arca built its own agent harness, persistent memory system, and deterministic safety guardrails around frontier models like Claude to safely handle client data and propose financial plans.

Despite Meta claiming that millions of users are shopping with its agentic AI tool Muse, executives across roughly 50 companies report that actual transactions and subscription cancellations driven by AI agents remain minimal. Even so, businesses like Chipotle and Costco are starting to optimize their websites for AI navigation, while others weigh the security risks and transaction fees associated with agentic commerce.

Wabi, an AI startup founded by Replika creator Eugenia Kuyda, has raised $55 million at a $1.25 billion valuation in a funding round led by IVP. The company recently pivoted from a prompt-based app builder to an AI messenger and personal agent that generates custom mini-apps for everyday tasks.

Deepgram is acquiring Daily, the company behind the open-source voice and video AI framework Pipecat, for an undisclosed amount. The two companies pledged to keep Pipecat open-source and vendor-neutral while also launching the Real-Time Context Standard (RCS), an open protocol for voice agents.

Amazon has added over 100 new shopping features to its Alexa assistant, enabling users to manage subscriptions, edit shopping lists, update delivery instructions, and resolve billing issues. The company also reported an eightfold increase in the use of its agentic features since the beginning of the year, with over 400 million consumers using Alexa to shop in the past 12 months.

An investigation by Flamerix AI found that Claude Code sends the full transcript of each background agent to its model every 30 seconds to generate a brief status line, consuming 15% to 20% of total token spend. Because there is no official setting to disable this default behavior, the authors released a Bun preload script that intercepts and disables the summary timer.

Latent Space interviews Supabase founders Paul Copplestone and Ant Wilson on how the platform is adapting its database and backend infrastructure for AI agents. The discussion covers recent announcements including Docker-free local development, declarative Git-based schemas, co-located compute, and the acquisition of Turso to scale for agent-generated applications.

A Wired journalist experimented with using multiple AI agents to manage a Premier League fantasy soccer team, finding that while multi-agent setups improved suggestions, the system still failed to beat the average human score. Researchers testing frontier models on prediction markets and benchmarks like Prophet Arena and Prediction Arena similarly note that AI forecasting is improving rapidly, though models still generally lag behind top human superforecasters and lose money in live trading.