AI Agent Analysis and Commentary
Independent analysis of AI agents and the models behind them from well-known writers, summarised with links.
-
Claude Mem hits 155.3K weekly downloads, 257% above its recent average
Claude Mem, a memory plugin for Claude Code, reached 155.3K npm downloads in the week ending Nov 19, 2026. That is more than double its previous weekly record of 75.3K.
-
Mastra sets a weekly download record, 36% above its 4-week average
Mastra, a TypeScript framework for building agents and AI workflows, reached 4M npm downloads in the week ending Nov 18, 2026. That is its highest weekly total in 75 weeks of data.
-
Fin Voice degraded for 1 h 35 min, 2.1x the typical incident
Fin, an AI customer agent for support chat, email and voice, had a minor Fin Voice incident on Nov 18. It ran from 16:27 UTC to 18:02 UTC, about twice as long as Fin's typical incident.
-
Skyvern outage lasted 12 h 5 min, 6.2x its typical incident
Skyvern Cloud had a major incident from Nov 17, 12:22 UTC to Nov 18, 00:28 UTC, during which browser sessions and agents failed to start. It is the longest Skyvern incident in the past 12 months.
-
Synthflow ElevenLabs voice incident ran 5 h 32 min, longest in 12 months
Synthflow, a platform for AI voice agents that answer and place phone calls, had a minor incident on Nov 18. Calls and API requests degraded for agents using ElevenLabs voices, and it lasted 13.3x the typical incident.
-
Droid's Gemini models were down for 39 h 51 min, 58.2x its typical incident
Factory's Droid, a software-engineering agent, lost access to its Gemini models from Nov 15, 04:01 UTC until Nov 16, 19:52 UTC. The incident was rated major and is the longest in the past 12 months.
-
[AINews] Gemini 4 Neon: GDM's Opus/Sol Contender
Google released Gemini 4 Neon, its middle-tier model, which scored 61 on the Intelligence Index. Anthropic priced its IPO at $83.
-
llm-gemini 0.38
llm-gemini 0.38 released following Google's introduction of Gemini 4 Neon model.
-
[AINews] GPT-6.2 Astra, the #2 Frontier Model is better and 20%-600% cheaper, but has mixed reviews; 6.2 Sol likely main workhorse
OpenAI released GPT-6.2 Astra and Sol models with lower benchmark costs and pricing. Early testers reported mixed experiences with Astra.
-
External sandboxes for coding agents
Simon Willison explores sandboxing approaches for coding agents to improve safety when running local models, in collaboration with Prime Radiant.
-
Fin outbound email delays lasted 21 h 20 min, 28.6x the typical incident
Fin, an AI customer agent for support conversations, had a major incident with delays in outbound email delivery. According to the official status page, it ran from Nov 12, 21:17 UTC to Nov 13, 18:36 UTC.
-
Corsair reaches 53.9K weekly downloads, up 70% on its 4-week average
Corsair, a tool for connecting your users to their apps, logged 53.9K downloads on npm in the week ending Nov 15, 2026. The previous weekly record was 33.3K.
-
Claude Cowork's Microsoft 365 connector errors lasted 27 h 57 min
Claude Cowork had elevated errors on its Microsoft 365 connector for 27 h 57 min, according to the official status page. The incident was rated major and ran 29.2x longer than the product's typical incident.
-
Novita Sandbox hits 132.1K weekly downloads, 351% above its average
Novita Sandbox, an SDK for code execution and cloud environments from novita.ai, logged 132.1K downloads in the week ending Nov 15, 2026. That is a new weekly record, well above its recent run rate.
-
Hebbia Chat requests failed for 1 h 15 min in a critical incident
Hebbia, an AI platform that analyses large document sets for finance and legal teams, had a critical incident on Nov 13 in which Chat requests were failing. According to the official status page, it lasted 1 h 15 min.
-
datasette-agent 0.6a0
Datasette Agent 0.6a0 released with bug fixes and adds an agent prompt to the Datasette homepage with table and query mentions.
-
Qwen Code hits a record 173.6K weekly downloads, up 70%
Qwen Code, Alibaba's open-source terminal coding agent, reached 173.6K downloads in the week ending Nov 10, 2026. That is a new weekly record and 70% above its recent average.
-
E2B template builds failed for 3 h 11 min, 2.2x its typical incident
E2B, which runs AI agent code in isolated cloud sandboxes, had template builds fail on Ubuntu package installs on Nov 11. The incident lasted 3 h 11 min, about 2.2x the product's typical incident.
-
Stagehand hits record 3.4M weekly downloads, 31% above its 4-week average
Stagehand, Browserbase's open-source SDK for browser agents, set a new weekly download record in the week ending Nov 9, 2026. It passed the previous high of 3M and now leads its category.
-
Deep Agents hits a record 2.9M weekly downloads, up 33%
Deep Agents, LangChain's agent harness built on LangGraph, reached 2.9M downloads in the week ending Nov 9, 2026.
-
AI Engineer Code Summit: OpenAI on New Ways of Working With Codex
OpenAI demonstrated new developer workflows for Codex including voice steering, cloud handoff, and dot notation features.
-
Designing for Agents — Zeno Rocha, Resend
Discussion on designing for AI agents, focusing on how Resend optimizes its documentation and tools for agent integration with Claude Code.
-
Introducing Our Smallest, Most Efficient Model: Muse Ember
Introducing Our Smallest, Most Efficient Model: Muse Ember Meta AI just launched their entry in the Claude Haiku/GPT Luna category: Muse Ember, which they describe as: [...] our smallest and most…
-
Week 46: 302.2M downloads across 492 agents, incidents down to 34
Downloads across the 492 AI agents we track rose +1% to 302.2M in the week ending Nov 9, 2026. Status page incidents fell to 34 from 57 the week before.
-
Amp hits a weekly download record of 66.2K, up 105% on its 4-week average
Amp, a coding agent from a Sourcegraph spin-out, logged 66.2K downloads in the week ending Nov 8, 2026. That beats its previous weekly record of 41.2K and is 105% above its recent average.
-
Metaharness downloads reach 156.9K in a week, 142% above average
Metaharness, a framework for scaffolding custom AI agent harnesses, logged 156.9K downloads in the week ending Nov 8, 2026. The previous weekly record was 82.6K, and a new release landed inside the same week.
-
Hide your slop PR descriptions behind
Discussion of PR descriptions generated by coding agents like Claude Code and Codex, noting they produce verbose descriptions that may obscure rather than clarify work.
-
ChatGPT Work errors lasted 13 h 7 min, its longest incident in 12 months
ChatGPT Work, ChatGPT's agent mode for long multi-step tasks, had an incident titled "ChatGPT Work and Dots Errors" from Nov 5, 15:25 UTC to Nov 6, 04:31 UTC. The official status page rates it minor.
-
Context Mode hits 59.1K weekly downloads, up 112% on its 4-week average
Context Mode, an MCP server for AI coding agents, set a weekly download record in the week ending Nov 5, 2026. The previous best was 36.4K, and the prior four-week average was 27.8K.
-
Codesign Mcp reaches 58K weekly downloads, 149% above its recent average
Codesign Mcp, an MCP design engine from img.ly, recorded 58K downloads in the week ending Nov 5, 2026. The previous weekly record was 42.2K, and the 4-week average before this week was 23.3K.
-
[AINews] Inkling 2 Small: 17A-594B trained in 68 hours on 12T tokens, beating Muse Spark
November 5, 2026 — Inside Inkling 2 Small’s training run and METR’s study of AI research at Anthropic.
-
datasette-apps 0.3a0
datasette-apps 0.3a0 improves LLM integration for creating custom HTML applications in Datasette, with app tagging support in chat.
-
datasette-agent 0.5a0
datasette-agent 0.5a0 adds a panel feature enabling the agent chat to sit beside Datasette pages for real-time app modification.
-
[AINews] LLMs can read & write neural network weights: "Weight Space Programming" cuts Autoresearch out of the RSI loop
Research demonstrates LLMs can directly manipulate neural network weights, with applications to reducing reasoning loops in models like Qwen and DeepSeek's agent prototypes.
-
Building AI for Lawyers - Jacob Lauritzen, CTO of Legora
Legora's CTO discusses building AI agents for lawyers, covering evaluation methods, large-scale document review, and challenges to the billable hour model.
-
Agents Prompt People: Using Agents for Long-Horizon Financial Work
Arca's head of AI discusses using agents for long-horizon financial work, including persistent memory and custom harnesses.
-
Using Agents for Long-Horizon Financial Work — Aman Singh, Arca
Arca's head of AI discusses the company's architecture for building an AI-native wealth management firm, including custom harness and persistent memory.
-
Why Agents Love Supabase
Podcast episode discussing Supabase as a backend platform for AI agents.
-
Coding From the Himalayas With an Offline AI Agent
Guide on setting up and running an offline coding agent using OpenCode, Ollama and Qwen.
-
GPT-6.1 Terra
OpenAI announced GPT-6.1 Terra, a model with visual reasoning and computer use capabilities alongside improved writing.