“Harrison Chase shared how Deep Agents’ summarization middleware reduces model-visible context without erasing work history, using a real or virtual filesystem to offload large tool results and older messages before previewing or summarizing them.”

#4 𝕏 Harrison Chase shared how Deep Agents’ summarization middleware reduces model-visible context without erasing work history, using a real or virtual filesystem to offload large tool results and older messages before previewing or summarizing them. Agents can also proactively trigger the same offload-and-summarize process with the compact_conversation tool.

“Harrison Chase said agent workspaces should be durable, inspectable, and swappable, highlighting MongoDB support for LangChain’s Deep Agents virtual file system.”

Harrison Chase said agent workspaces should be durable, inspectable, and swappable, highlighting MongoDB support for LangChain’s Deep Agents virtual file system. Its BackendProtocol lets agent code use read, write, glob, and grep operations while teams choose their production storage layer.

“#7 𝕏 Harrison Chase shared a free 3-hour YouTube Deep Agents course that dives into task planning, file-system-based context management, subagent spawning, and long-term memory techniques.”

#7 𝕏 Harrison Chase shared a free 3-hour YouTube Deep Agents course that dives into task planning, file-system-based context management, subagent spawning, and long-term memory techniques. #15 𝕏 Harrison Chase enabled cache-aware requests in Deep Agents, reusing a warm cache to slash cache misses and drive down operational costs.

“𝕏 Harrison Chase (@LangChain) spotlights a community guide on building a Claude Code–style agent using LangChain’s Deep Agents framework.”

#4 𝕏 Harrison Chase (@LangChain) spotlights a community guide on building a Claude Code–style agent using LangChain’s Deep Agents framework. He highlights how leveraging the strong performance of GLM-5.2 can boost your custom agent’s capabilities.

“Harrison Chase reports that open-source models now excel at file operations, summarization, tool use, and retrieval—capabilities strong enough to power agent frameworks like Deep Agents.”

#16 𝕏 Harrison Chase reports that open-source models now excel at file operations, summarization, tool use, and retrieval—capabilities strong enough to power agent frameworks like Deep Agents. #17 𝕏 Mustafa Suleyman launched three Microsoft AI models in just months: MAI-Transcribe-1 (world’s most accurate transcription across 25 languages per FLEURS WER), MAI-Voice-1 for ultra-natural speech, and MAI-Image-2 (top 3 model on Arena).

“Harrison Chase reports Jensen Huang’s Interrupt fireside on enterprise agents, unveiling a LangChain x NVIDIA partnership and launching Deep Agents powered by Nemotron models via the NVIDIA Agent Toolkit.”

Today's top 25 insights for PM Builders, ranked by relevance from X, LinkedIn, YouTube, and Blogs. Alibaba Launches Qwen3.5-Omni: Builds Websites From Video #1 𝕏 Qwen unveiled Qwen3.5-Omni, a native omni-modal AGI that understands text, image, audio and video and features “Audio-Visual Vibe Coding” to instantly build websites or games from a vision prompt. Offline it offers script-level captioning, outperforms Gemini-3. #2 in Dharmesh Shah reports that OpenAI has launched Codex support for Claude Code—extending ChatGPT subscriptions into JetBrains, Xcode, OpenCode, Pi and more. #3 𝕏 Claude launched “Claude Code,” letting the AI open your apps, navigate UIs, and test what it built—all from the CLI. It’s now in research preview on Pro and Max plans. #4 𝕏 Harrison Chase reports Jensen Huang’s Interrupt fireside on enterprise agents, unveiling a LangChain x NVIDIA partnership and launching Deep Agents powered by Nemotron models via the NVIDIA Agent Toolkit. #5 𝕏 Guillermo Rauch launched Opus 4.5, ushering in agent-driven coding, and shared early “agenting responsibly” guidance to temper LLM overconfidence while prioritizing security, durability, and availability. #6 𝕏 Harrison Chase rebuilt LangChain’s GTM agent on Deep Agents and DeeplineCLI, automating lead enrichment, outreach, and conversion workflows. #7 𝕏 Teresa Torres adds a PreToolCall hook on ExitPlanMode to block its default tool call and trigger her custom plan skill instead. #8 𝕏 Teresa Torres reports that Zapier’s core automation has degraded—zaps often fail—and she now asks Claude to build a custom webhook listener for more reliable triggers and error handling. She’s also moving off Airtable due to similar quality issues. #9 𝕏 Santiago unveils Pokee_AI’s zero-setup agent platform—instant signup access to sandboxed AI execution with role-based access control, encrypted credential vaults, long context memory, and 70% lower token consumption than OpenClaw. #10 𝕏 claire vo 🖤 launched “Gridley’s Anti-System for Automating Life with Claude” and shared a full step-by-step guide. Find the detailed walkthrough on the @chatprd AI blog. #11 ▶️ How to turn Claude code into your personal life operating system | Hilary Gridley How I AI Podcast Configuring Claude Code in the macOS terminal to automate life admin by capturing to-dos via an iPhone back-tap shortcut, storing context in local markdown files, and running a custom “plan my day” workflow that schedules events to Google Calendar and logs daily activities. The iPhone shortcut uses Apple Shortcuts’ “Dictate Text” action triggered by Accessibility > Touch > Back Tap > Double Tap to append spoken items (e.g., “reschedule pediatrician appointment”) into a reminders inbox markdown file. Claude Code is installed by copying the install line from the Claude docs into the terminal, then launched with the “claude” command to read and edit context files (e.g., reminders.md, preferences.md) in a dedicated folder. The “plan my day” Claude Code command pulls tasks from reminders.md, scheduling preferences learned in preferences.md (e.g., pumping windows, childcare), and existing Google Calendar events, then creates new 🦛-tagged calendar slots (e.g., a 10-minute “make post office appointment” for a baby passport) and writes a daily note comparing planned vs actual tasks. #12 ▶️ Stop Vibe Coding. Start Getting Customers. Greg Isenberg Greg Isenberg outlines seven distribution strategies for AI-built products, including using the OpenAI MCP protocol to build MCP servers that achieved 150+ installations in 30 days with zero ad spend, leveraging programmatic SEO to spin up 10,000 pages in 48 hours, and acquiring niche newsletters for $5,000–$20,000. 200,000 new vibe coding projects are launched daily on Lovable An MCP server built via the OpenAI MCP protocol secured over 150 installations in 30 days at $0 ad spend in a fintech use case A 10,000-subscriber niche newsletter can be purchased for $5,000–$20,000 through platforms like Deuce.com #13 𝕏 clem 🤗 warns that inadequate tooling and poor fine-tuning—not the capacity of smaller local models—are behind most deployment failures. #14 📝 Simon Willison Georgi Gerganov on why it's hard to find local models that work well with coding agents - Georgi Gerganov explains that the main problems with local models stem from fragility across a long chain of components (harness, chat templates, prompts, inference) developed by different parties, making reliable behavior difficult to achieve. Even if individual pieces seem to work, subtle breakages can exist elsewhere in the stack. #15 in Colin Matthews reveals that AI agents actually don’t retain memory beyond each prompt’s context window and can be built without specialized frameworks by simply looping LLM API calls. #16 in e Carl Vellotti demos the full Claude Code OS in his third deep-dive with Aakash Gupta, after the first two episodes crossed 1M+ views. #17 𝕏 Ali Ghodsi echoes Jeff Dean that legacy, human-paced tools bottleneck AI agents. He introduces Lakebase Postgres, offering instant branching, snapshots, and sub-second auto-scaling—orders of magnitude faster than traditional databases. #18 📝 Doug Turnbull Stop evaluating search with queries - Doug argues that traditional query-based evaluation of search is flawed and recommends using judgment lists and transformed clickstream data to produce more reliable evaluation labels. This approach better captures result relevance than treating queries as the sole evaluation unit. #19 𝕏 clem 🤗 argues that as no-code tools make app building ubiquitous, true differentiation comes from training, optimizing and running your own AI models. #20 in Peter Yang highlights how Jenny, Claude’s design lead, uses Cowork to auto-summarize user feedback into a weekly product-priorities deck shared via Slack and maintains a simple folder-based “memory system” to keep Claude’s outputs up to date. #21 𝕏 claire vo 🖤 dives into how @yourgirlhils scripts Claude Code to build a personal productivity OS—automating tasks, managing routines, and prepping meetings—in a 52-minute deep dive. #22 𝕏 Lenny Rachitsky highlights Claire Vo’s "Sage," an OpenClaw-powered bot that automates project management and weekly LinkedIn reminders for her Maven course. It keeps her on track for launch without the need to hire ops or marketing staff. #23 𝕏 There's An AI For That launched SureThing, an AI agent that remembers your voice, goals and workflows and acts across 1,000+ apps. It features persistent memory that sharpens over time and serves as a cloud-first OpenClaw alternative. #24 𝕏 Peter Yang confirms that @cursor_ai works flawlessly in China with every model type. #25 𝕏 Qwen demos a fresh Audio-Visual Vibe Coding system, turning sound inputs into synchronized visual effects in real time. Found this valuable? Share it with another PM - they can subscribe at genaipm.com Unsubscribe • Switch to Weekly