Claude Code Gets Agent View: One Dashboard for All Your AI Sessions

🧠 LAUNCH

Claude Code Gets Agent View: One Dashboard for All Your AI Sessions

The tmux for AI coding agents is here. Claude Code ships agent view — a single dashboard to monitor, manage, and switch between all running AI sessions in real time. If you've been juggling multiple terminal tabs with separate Claude instances, this collapses that workflow into one pane of glass. It shipped today as a research preview in v2.1.139. Run claude agents to try it. (10,169 likes | 640 RTs) Read more →

Anthropic Launches Native Claude API on AWS — No Bedrock Required

Claude Platform on AWS breaks the Bedrock-only constraint. AWS customers now get the full Messages API, Files API, Managed Agents, and code execution — all with native AWS billing and IAM auth. For enterprise teams that wanted Claude's latest capabilities but were stuck waiting for Bedrock support, the bottleneck is gone. Read more →

OpenAI Launches Daybreak to Put Frontier Models on Cyber Defense. Daybreak combines frontier models with Codex and security partners to accelerate cyber defense — arriving days after Anthropic's Mythos pentesting results. Frontier labs are racing to own the cybersecurity vertical, and the timing isn't accidental. (4,630 likes | 458 RTs) Read more →

Qwen Ships WebWorld: Open Models Purpose-Built for Web Agents. WebWorld drops in 8B/14B/32B sizes, Apache 2.0 licensed, with +9.9% on MiniWob++ and +10% on WebArena over prior open models. The first serious open-source competitor for web automation tasks that were previously locked behind proprietary APIs. (441 likes | 69 RTs) Read more →

MiniCPM-V-4.6 Continues the Compact Multimodal Push. OpenBMB drops the latest MiniCPM-V, continuing the push for high-quality vision-language models small enough to run locally. The MiniCPM series has consistently punched above its weight class on multimodal benchmarks — worth testing for on-device use cases. (179 likes) Read more →

🔧 TOOL

Claude Code v2.1.139: /goal Turns Single Sessions into Autonomous Loops

Beyond agent view, this release adds /goal — set a completion condition and Claude keeps working across turns until it's met. Works in interactive mode, the -p flag, and Remote Control. Combined with agent view, this turns Claude Code from a coding assistant into a genuine multi-agent orchestrator. Update now. Read more →

Codex Gets a Native OpenAI Plugin for Building AI Apps. Codex can now use OpenAI APIs directly through a built-in plugin — removing the copy-paste loop between docs and coding. A practical productivity bump for anyone building on the OpenAI stack. (1,580 likes | 109 RTs) Read more →

Anthropic Python SDK v0.101.0 Adds Native AWS Client. The SDK-level support that makes today's Claude Platform on AWS launch immediately usable in production code. If you're evaluating the new AWS integration, start here. Read more →

📝 TECHNIQUE

Karpathy Validates HTML as the New LLM Output Format

Ask your LLM to structure responses as HTML — slideshows, visualizations, structured docs. Karpathy validates the pattern with 9.4K likes, building on swyx's "HTML is the new markdown" thesis from last week. The key insight: LLMs already know HTML deeply, so you get rich formatting for free without custom rendering logic. Try adding "structure your response as HTML" to your next prompt. (9,452 likes | 894 RTs) Read more →

Simon Willison Puts LLMs in a Unix Shebang Line. Willison figured out how to use his LLM CLI tool in a shebang — meaning you can write executable scripts in plain English or hook up YAML templates. #!/usr/bin/env llm at the top of a file, natural language below, executable output. A small hack with big implications for scripting workflows. (233 likes | 12 RTs) Read more →

Agentic Search Is 80% of Context Engineering. Leonie's AI Engineer workshop argues that most agent failures are retrieval failures, not reasoning failures. The full tool landscape — shell tools, semantic search, query execution, agent skills — gets covered with guidance on when each approach breaks. Worth watching if you're building agents that need to find their own context. (237 likes | 36 RTs) Read more →

🔬 RESEARCH

Mollick Flags AI's Creative Convergence Problem. Ethan Mollick highlights a fundamental limitation: AI models converge on similar ideas, limiting their utility for science and creative work. A new paper shows you can optimize models specifically for creativity — a capability dimension the major labs haven't prioritized yet. If your use case requires genuine novelty, this is worth reading. (357 likes | 44 RTs) Read more →

Lilian Weng Reveals the 137-Page Logbook Behind a Training Run. OpenAI's Lilian Weng shares that a recent training run involved 12 versions and 137 pages of documentation. The takeaway: frontier model training is as much human coordination as it is compute — and the craft of iteration, documentation, and decision-making is what separates successful runs from expensive failures. (268 likes | 16 RTs) Read more →

💡 INSIGHT

OpenAI Creates a Deployment Subsidiary to Win the Enterprise Last Mile

OpenAI creates a majority-owned subsidiary — the "Deployment Company" — bringing together 19 investment firms and consultancies to help enterprises deploy AI. This is OpenAI pivoting from model provider to full-service deployment partner, a strategic bet that API revenue alone won't capture enterprise value. The move signals that the frontier AI business model is shifting from "best model wins" to "best integration partner wins." (8,073 likes | 956 RTs) Read more →

Google Confirms Criminal Hackers Used AI to Find a Major Vulnerability. The NYT reports that criminal hackers used AI to discover a major software flaw — crossing from theoretical risk to confirmed real-world exploitation. This directly validates the urgency behind both OpenAI's Daybreak launch and Anthropic's Mythos security work this week. Brief your security team. (96 likes | 68 RTs) Read more →

Gemini Flash 3.2 All But Confirmed for Google I/O. Bindureddy says Flash 3.2 is essentially confirmed for I/O, and they've already replaced GPT 5.5 low with Flash in 70% of scheduled jobs. If Flash 3.2 can match GPT 5.5 low at a fraction of the cost, the high-volume inference pricing equation shifts dramatically. (217 likes | 5 RTs) Read more →

🏗️ BUILD

Training an LLM in Swift: From Gigaflops to Teraflops. A deep technical walkthrough of implementing LLM training in Swift, optimizing matrix multiplication from gigaflops to teraflops on Apple silicon. Practical systems-level knowledge for anyone curious about what's actually happening beneath the PyTorch abstractions — and a compelling case that Swift's performance ceiling is higher than most ML engineers assume. (209 likes | 11 RTs) Read more →

🎓 MODEL LITERACY

Context Engineering: Today's agent view launch and agentic search coverage both tackle the same core problem: routing the right information into the right context window at the right time. Context engineering — structuring, selecting, and compressing information for LLM consumption — is why multi-session agents need dashboards and why most agent failures are retrieval failures, not reasoning failures. A model with perfect reasoning but the wrong context will confidently produce the wrong answer. When you hear "context window," think less about the size limit and more about what goes into it — that curation step is where most of the real engineering happens.

⚡ QUICK LINKS

Agent view announcement: Claude Code's most engaged feature launch hits 10K likes on X. Link
Claude's Constitution as audiobook: Narrated by its authors Amanda Askell and Joe Carlsmith. (1,907 likes | 224 RTs) Link
Consensus raises $30M: Building the AI OS for research with 2.5M researchers already on the platform. (207 likes | 55 RTs) Link
Capital, not compute: The Information argues financing is now the binding constraint on AI infrastructure. Link
LangChain Core 1.4.0: Major milestone with Pydantic v1 fixes and dependency updates. Link
Interfaze: A new transformer variant claims improved accuracy-at-scale characteristics. (100 likes | 27 RTs) Link

🎯 PICK OF THE DAY

OpenAI creating a majority-owned deployment subsidiary signals that API revenue alone can't capture enterprise value. The "Deployment Company" isn't an acqui-hire or a consulting arm — it's 19 investment firms and consultancies organized under OpenAI's majority ownership to handle the last mile of enterprise AI integration. The real margin in enterprise software has always lived in implementation, not licensing, and OpenAI just said the quiet part out loud: the best model doesn't win, the best deployment partner does. This reshapes the competitive landscape fundamentally. Anthropic, Google, and Meta are competing on model quality; OpenAI is now competing on who gets the contract to rewire your business. For builders, the implication is clear — if the model layer commoditizes (and the pace of catch-up benchmarks suggests it will), the moat moves to integration, tooling, and trust. Watch whether Anthropic responds with its own services play or doubles down on the platform-and-partners model. (8,073 likes | 956 RTs) Read more →

Until next time ✌️