News
The latest from the AI agent ecosystem, updated multiple times daily.
Sem swaps line diffs for function diffs, claims agents read them 2.3x better
Ataraxy Labs' sem is a single binary that gives git semantic awareness: diff, blame, impact and log at the level of functions and classes across 26 languages. Its boldest claim is that agents are 2.3 times more accurate on sem output than on raw line diffs.
Universal Memory Protocol pitches itself as the MCP of agent memory
UMP is a transport-neutral spec for portable agent memory: signed, bi-temporal records that any harness can read and any store can serve. Its pitch is the third interoperability layer, after MCP for tools and A2A for agent-to-agent coordination.
Agentic coding's token bill: 59% goes to review, not writing code
A new arXiv study traced every token through 30 software tasks run on ChatDev with a GPT-5 reasoning model. Code review consumed 59.4% of tokens on average, making verification, not generation, the dominant cost of agentic software engineering.
Jane Street's designer ships 2,000-line prototypes instead of Figma mockups
Designer Edwin Morris says his Figma use has fallen off a cliff since Claude let him build working prototypes directly in Jane Street's OCaml codebase. The catch: engineers now review fully baked features instead of shaping proposals.
Ten years of payments expertise, reduced to a prompt
A backend engineer with a decade in payments and finance describes watching three pillars of his expertise erode in sequence: domain knowledge, debugging, then architecture. The detail that stings most is that his employer no longer hires by specialisation at all.
Meta's chatbot hack and OpenAI's Lockdown Mode are the same story
In the same week, Meta confirmed more than 20,000 Instagram takeovers carried out through its AI support chatbot, and OpenAI shipped a mode that amputates ChatGPT's riskiest capabilities. Together they show an industry quietly giving up on preventing agent misuse and engineering for blast radius instead.
Gemma 4 gets a sub-1GB build that runs on a phone
Google has released quantization-aware-trained checkpoints for Gemma 4, shrinking the E2B text model to under 1GB of memory. A custom mobile format and selective 2-bit compression keep quality close to the full-precision reference.
Microsoft puts durable execution inside Postgres, no extra service
Microsoft has open-sourced pg_durable, a PostgreSQL extension that runs long-running, fault-tolerant workflows entirely inside the database. It checkpoints each step, so a crash resumes from the last good point instead of forcing you to rebuild state.
Sakana AI formalises its bet on self-improving AI with a dedicated RSI Lab
Sakana AI has established a Recursive Self-Improvement Lab tasked with redesigning the AI development process with AI. Its pitch is sample efficiency: self-improvement that compounds on national rather than hyperscale compute budgets.
Alibaba open-sources the code reviewer it ran internally for two years
Alibaba has released Open Code Review, the AI review tool it says served tens of thousands of its own engineers and flagged millions of defects. It pairs deterministic rule pipelines with an LLM agent that reads the whole codebase, not just the diff.
ChatGPT's Lockdown Mode reaches everyone, free tier included
OpenAI is rolling out Lockdown Mode to all personal ChatGPT accounts and self-serve Business plans. The setting trades agentic features like browsing and agent mode for hard guarantees against prompt-injection data theft.
Meta's AI chatbot reset Instagram passwords for anyone who asked
Meta has notified more than 20,000 people that their Instagram accounts were hijacked through its AI chatbot. A flaw let attackers ask the bot to send password reset links to email addresses they controlled.
Anthropic open-sources its vulnerability-hunting harness for Claude
Anthropic has released the Defending Code Reference Harness, an open-source blueprint for pointing Claude at a codebase to find and patch security bugs. It ships an autonomous scanner and a customise skill, and is candid about where the approach falls short.
Google will pay SpaceX $920m a month for GPUs it says it suddenly needs
Google has agreed to pay SpaceX US$920 million per month from October 2026 to June 2029 for access to roughly 110,000 Nvidia GPUs. The company calls it bridge capacity for unexpected Gemini Enterprise demand.
S&P 500 refuses to bend for SpaceX, closing the fast lane for OpenAI and Anthropic
S&P Dow Jones Indices has rejected rule changes that would have fast-tracked SpaceX into the S&P 500 after its IPO. The same waivers were the only quick route in for OpenAI and Anthropic, both still unprofitable.
Anthropic says 80% of its merged code is now Claude's
Anthropic's research institute published internal data showing AI is already accelerating AI development, and set out what a credible global pause would demand. The standout figure: more than 80% of the code merged into Anthropic's own codebase is now written by Claude.
"MCP is dead" keeps killing the wrong thing
The MCP obituaries have the receipts on context bloat. They also conflate a calling convention with a protocol, and the protocol's own author shipped the fix while the standard got donated to a foundation. The angle: what is actually dying is loading every tool you own into a window you pay for, not interoperability itself.
Someone finally charted the rsync AI-bugs panic. The data says no
A distributional analysis of 37 rsync releases finds the two with Claude-assisted commits sit squarely in the middle of the project's historical bug rate, not the tail. The worst release on record had no AI involvement at all, and nobody complained.
Distilling multi-agent debate into one model cuts tokens by up to 93%
A new paper folds multi-agent debate into a single LLM through fine-tuning, matching or beating the full debate while using up to 93% fewer tokens. The internalised agents show up as separate, steerable directions in the model's activation space.
Claude now writes most of Anthropic's code, and Anthropic wants a pause button
The Anthropic Institute says more than 80% of code merged into its production codebase in May 2026 was authored by Claude, and engineers now ship 8x as much code per quarter as in 2024. The piece argues recursive self-improvement is not here yet but could arrive sooner than institutions are ready for.
Alibaba open-sources the code reviewer it ran internally for two years
Alibaba has released Open Code Review, the AI review tool it used internally across tens of thousands of developers. It pairs deterministic pipelines with an LLM agent to fix the two failures of general-purpose review agents: skipped files and wrong line numbers.
The numbers say Claude did not break rsync
After a viral post blamed Claude-assisted commits for regressions in rsync, an independent analysis ran the bug data across every release. The verdict: the two Claude releases are statistically indistinguishable from history. The outrage rested on a single tail event.
Microsoft puts durable workflow execution inside Postgres itself
Microsoft has open-sourced pg_durable, a Postgres extension that runs crash-resilient workflows entirely inside the database with no external orchestrator. A workflow is a graph of SQL steps that checkpoints as it goes and resumes from the last good point after a crash.
Microsoft puts durable execution inside Postgres itself
Microsoft has open-sourced pg_durable, an extension that runs Temporal-style durable workflows inside PostgreSQL with no extra service. You define the workflow as a graph of SQL steps and the database checkpoints each one, resuming after a crash. It ships inside Microsoft's new Azure HorizonDB.
Alibaba open-sources the code reviewer it ran internally for two years
Alibaba has released Open Code Review, the AI reviewer it says served tens of thousands of its own engineers and flagged millions of defects. It pairs deterministic rule pipelines with an LLM agent that can read the whole codebase, not just the diff.
Anthropic open-sources the harness behind its vulnerability-hunting agent
Anthropic has published the Defending Code Reference Harness, a reference build of the autonomous agent it uses to find, verify and patch software vulnerabilities. It runs Claude through a full recon-to-patch loop and refuses to operate outside a gVisor sandbox.
Anthropic open-sources the loop behind its Claude security scanner
Anthropic has released a reference implementation of the autonomous pipeline it uses to find and patch code vulnerabilities with Claude. It is the open version of the recon-to-patch loop behind Claude Security and the Mythos preview. The catch: the part that actually hunts memory bugs refuses to run outside a sandbox.
Cognition and Cursor are pricing opposite bets on the same assumption
Cognition just raised over $1 billion at a $26 billion valuation for its autonomous agent Devin. Cursor is reportedly raising at $50 billion for the opposite theory of how coding agents win. Both numbers rest on the same thing being true, that the company between the developer and the model keeps the margin, and Anthropic's Claude Code is the reason it might not.
AI Can Find the Bug. Verifying It Is Still the Whole Job
A controlled experiment turned a dozen frontier models loose on a deliberately vulnerable app; most scored zero and only GPT-5.5 cleared it reliably. Read alongside the AI slop that killed curl's bug bounty and AISLE's 12-of-12 CVE run on OpenSSL, the lesson isn't whether agents can hack. Discovery got cheap this year, verification didn't, and that gap is where the economics of agentic security actually break.
Two coding agents, one git repo: a tiny protocol lets Claude Code and Codex talk
A new feature in h5i, an 'AI-aware' Git, lets Claude Code and Codex hand work back and forth by writing messages into the repository itself. No server, no socket. Each message is one JSON line on a dedicated git ref, so the whole conversation is versioned and merges without conflicts.
YC's Hyper bets the missing piece for AI teams is shared context
Hyper, a Y Combinator startup, launched a "company brain" that ingests a team's activity across its tools and injects the resulting context into every AI chat turn. The pitch: today's models are capable but ignorant of your company, and that gap is the real bottleneck.
A $1,500 test of which LLMs will actually hack an app, and which refuse
Security researcher Kasra Rahjerdi built a deliberately vulnerable app and turned a field of models loose on it. GPT-5.5 solved it 7 of 10 times; DeepSeek V4 Pro was about 15x cheaper per success; Gemini 3.1 Pro refused to try. A scrappy test, not a benchmark.
Mathematicians draw a line as AI clears 52% of FrontierMath
The Leiden Declaration, backed by the International Mathematical Union, warns that AI could flood mathematics with plausible-but-flawed proofs and hand research priorities to tech firms. It lands as GPT-5.5 Pro tops the FrontierMath benchmark at 52.4%.
Ideogram open-weights a 9.3B image model that out-renders 32B rivals
Ideogram released 4.0, its first downloadable model: a 9.3B-parameter diffusion transformer with open weights. It claims better text rendering than models several times its size, and takes structured JSON prompts for precise layout control.
Anthropic's agent sandboxes held; its own proxy code didn't
Anthropic published how it contains Claude across claude.ai, Claude Code and Cowork, using a different isolation layer for each. Its blunt takeaway: the off-the-shelf sandboxing primitives held, while the custom code wrapped around them was where things broke.
Cloudflare buys VoidZero, putting Vite's toolchain behind its edge
Cloudflare has acquired VoidZero, the company Evan You founded to unify JavaScript tooling around Vite, Vitest, Rolldown and Oxc. The team joins Cloudflare's Emerging Technology group and the tools stay open source. Cloudflare is also seeding a $1M fund for Vite maintainers independent of both companies.
Uber caps engineers at $1,500 a month per AI coding tool
After running through its 2026 AI budget in four months, Uber is limiting each employee to $1,500 of monthly token spend per coding tool. The cap doubles as the clearest dollar signal yet for what agentic coding is worth to a big employer.
Gemma 4 12B drops the multimodal encoder entirely
Google's new 12B open model runs agentic multimodal workloads on a 16GB laptop, and it gets there by removing the separate image and audio encoders most multimodal models depend on.
Liquid AI's 24B MoE Runs on Your Laptop
Liquid AI releases LFM2-24B-A2B, a 24 billion parameter Mixture of Experts model with only 2.3 billion active parameters per token. The model fits in 32GB of RAM, making it deployable on consumer hardware including laptops with integrated GPUs and NPUs. It shows consistent quality gains on benchmarks like GPQA Diamond and MMLU-Pro as the LFM2 family scales from 350M to 24B parameters. Day-one support for llama.cpp, vLLM, and SGLang, with competitive throughput against Qwen3-30B-A3B and gpt-oss-20b.
Open-source DAC lets AI agents build dashboards humans can review
DAC is an open-source Dashboard-as-Code tool that lets you write dashboards in YAML or TSX. The key idea: it's built so AI agents can create dashboards that humans can actually review and approve. Ships with a Codex-powered AI agent for live updates, supports major databases through Bruin, and includes a semantic layer for reusable metrics and dimensions.
Claude Code Won't Read AGENTS.md, and That's a Problem
A GitHub feature request asks Claude Code to support AGENTS.md, the emerging standard file format for AI coding agents. Tools like Codex, Cursor, and GitHub Copilot already read it. Claude Code uses its own CLAUDE.md, forcing teams with multiple AI tools to maintain duplicate files.
UPenn's Codex skill renders web page videos from plain English
UPenn researchers released web-scroll-video, an open-source tool that records web pages as MP4s using headless Chrome and FFmpeg. Built as a skill for OpenAI's Codex, it lets you describe video actions in plain English and generates the video from those cues. The code is on GitHub under UPenn's CIS organization.
Governor cuts Claude Code token waste by 55%
Governor is a plugin for Claude Code that optimizes context usage and reduces token waste through compact professional output, context hygiene, tool-output filtering, and usage telemetry. It features memory compression, protected-span safety, quality guards, and planning guardrails for coding tasks.
SimplePDF's local AI copilot fills forms without phoning home
SimplePDF Copilot lets you fill PDF forms through conversation. The tool uses client-side tool calling with local models, so document data stays on your machine. Designed for embedded, white-labeled deployments in customer products.
SKILL.make: Agent Skills as Makefiles Cut Tokens 15%
Developers can now define AI agent skills using Makefile syntax. SKILL.make replaces prose with structured dependency graphs, cutting token usage roughly 15% in testing.
Rust via Claude: This Gopher Isn't Converting
A Go developer used Claude as a pair programmer to learn Rust by building a chat server, then compared the two languages on enums, error handling, async runtimes, and debugging tools.
DeepSeek V4: almost frontier, a fraction of the price
Simon Willison reviews DeepSeek's new V4 model series, featuring Pro (1.6T parameters, 49B active) and Flash (284B parameters, 13B active) models with 1M token context and MIT license. Both models offer dramatic cost advantages over frontier models from OpenAI, Anthropic, and Google. Flash is the cheapest small model at $0.14/M input, while Pro is the cheapest larger frontier model at $1.74/M input. Benchmark comparisons show competitive performance with much improved efficiency over DeepSeek V3.2.
Software Jobs Up 11% Even as AI Spending Hits $650B
Citadel Securities analysis challenges AI displacement narratives, showing software engineer job postings up 11% YoY despite $650 billion in AI capital expenditure. AI adoption follows S-curve patterns rather than exponential growth, with stable real-time data showing little evidence of imminent labor displacement. The wrinkle: companies want senior architects, not junior coders, as AI tools handle entry-level work.
The end of "Just ask Sarah"
Every team has a Sarah who holds the institutional knowledge. AI agents can't walk over and ask her. Simon Aronsson argues that as agents start writing code, documentation like ADRs and specs shifts from courtesy to necessity, because agents extend existing patterns without understanding the reasoning behind them.
Omar orchestrates 100 AI coding agents from your terminal
Omar is a terminal user interface (TUI) for creating and managing agentic organizations with deep hierarchies of parallel AI agents. Built on tmux, it lets you mix heterogeneous backends like Claude Code, Codex CLI, Cursor, and Opencode, with full control to navigate and interact with any subagent.