Opus 4.7 tops SWE-bench Pro, OpenAI Codex watches your screen, and GitHub trending is nothing but agent infrastructure.
> Everyone debated which model scored higher on SWE-bench. Meanwhile, the real shift happened underneath: agents stopped being demos and started becoming infrastructure. The model wars matter less when the agent layer is where developers actually live.
Anthropic released Claude Opus 4.7 on Wednesday, and the benchmarks are significant. SWE-bench Pro jumped from 53.4% (Opus 4.6) to 64.3%, surpassing OpenAI's GPT-5.4 at 57.7%. CursorBench hit 70%, up from 58%. Visual acuity nearly doubled to 98.5%, and image input resolution tripled to approximately 3.75 megapixels with a 2,576px long edge. On Rakuten-SWE-Bench, it resolves 3x more production tasks than its predecessor.
But the benchmarks aren't the story. The story is what Anthropic deliberately removed.
Under an internal initiative called "Project Glasswing," Anthropic reduced cybersecurity capabilities during training. The model ships with automatic safeguards that block prohibited security uses. If you're a legitimate penetration tester, you now need to apply for a Cyber Verification Program to unlock those capabilities. This is the first time a frontier lab has publicly shipped a model that is intentionally worse at a specific domain for safety reasons. Whether you think that's responsible or paternalistic depends on your threat model — but it's a precedent either way.
Pricing stayed flat at $5/$25 per million input/output tokens. The catch: a new tokenizer maps identical text to up to 35% more tokens, so your actual costs may increase despite unchanged per-token rates. A new xhigh effort level slots between high and max for compute-intensive tasks. The model follows instructions more literally than 4.6, which is a feature for structured workflows and a headache for prompts that relied on the model reading between the lines.
The HN thread hit 4,254 points and 1,261 comments. The top debate: whether deliberate capability reduction in one domain sets a precedent for future restrictions in others.
Migration advice: test your existing prompts before switching. The literal instruction-following will break workflows that depended on creative interpretation.
The Codex story deserves more than a table row. OpenAI didn't just add features to a coding tool — they repositioned it as a fundamentally different category. Background Computer Use gives the agent its own cursor on macOS, letting it see your screen, click, and type without interrupting your workflow. Multiple agents run concurrently. The agent can browse the web and accept inline comments on pages. Over 90 new plugins cover JIRA, GitLab, Microsoft Suite, and Slack.
The most provocative feature: autonomous task scheduling. Codex can wake itself up to continue working on projects across days or weeks. It's the first major product to treat an AI coding agent as a daemon, not a request-response tool.
Currently macOS-only, with EU/UK access delayed. It hit 877 HN points — strong, but notably cooler reception than Opus 4.7's 4,254. The community seems more impressed by raw capability gains than by new interaction paradigms. That might be a mistake.
China's National Security Commission branded Meta's $2B Manus acquisition as an attempt to "hollow out the country's technology base." Co-founders Xiao Hong and Ji Yichao — who relocated Manus from Beijing to Singapore in summer 2025 — have been summoned by China's NDRC and barred from leaving the country. A multi-agency review is underway using export controls, investment, and competition laws. Chinese investors are reportedly discussing unwinding the deal.
The precedent matters: China is retroactively blocking AI talent and IP transfers even after a company has legally relocated. If you're an AI startup with any Chinese founding team members, your M&A playbook just got significantly more complicated.
The hottest topic in AI engineering this week isn't a model release. It's memory.
Marco Somma's dev.to post "I Ran 500 More Agent Memory Experiments" crystallized something the community has been circling for months: the hard problem in agent memory isn't recall — it's binding. Your agent can retrieve relevant context from a vector store. What it can't do is figure out which memories belong to which task, which user, or which session when things get interleaved.
Think about it concretely. You have an agent that's been helping three developers on the same codebase. Developer A deployed a hotfix for the auth service last Tuesday. Developer B refactored that same service on Thursday. Developer C asks the agent to "continue where we left off on the auth service." The vector search returns five relevant memories. Three are from the correct developer's session. One is from Developer A's hotfix. One references a service name that was changed two months ago. The agent confidently merges context from all five.
This is the binding problem. Semantic similarity can't distinguish temporal relevance, task ownership, or session scope.
The GitHub trending page tells you the community is tackling this from multiple angles. The forrestchang/andrej-karpathy-skills repo — literally a single CLAUDE.md file — gained nearly 8,000 stars in a single day. It encodes behavioral rules that shape how Claude Code operates: what mistakes to avoid, what patterns to prefer, how to handle ambiguity. That's not RAG. That's not vector search. It's a prompt-level memory contract that solves binding by making scope explicit at definition time.
Then there's thedotmack/claude-mem at 60,791 stars, which captures full Claude Code sessions, compresses them with AI, and injects that compressed context into future sessions. The compression step is key — it's not just storing memories, it's curating which context carries forward and which gets dropped. And lsdefine/GenericAgent at 3,175 stars takes a different approach entirely: a self-evolving agent that grows a skill tree from a 3.3K-line seed, claiming 6x less token consumption than baseline approaches.
The pattern across all three: developers are building memory as explicit infrastructure, not as a model feature.
Somma's experiments showed something counterintuitive: agents with high recall but poor binding performed worse than agents with less total memory. Stale or misattributed context didn't just add noise — it actively poisoned decision-making. The agent would confidently apply a deployment pattern from three sprints ago because it was semantically similar, ignoring that the infrastructure had changed.
The practical takeaway for anyone building agent memory systems right now: stop optimizing your embedding model and start building explicit scope tags. Timestamp every memory. Tag it with the task ID, user, and session. Make deletion a first-class operation — strategic forgetting is as important as comprehensive recall. And test your memory system with interleaved multi-user scenarios, not single-user sequential benchmarks.
andrej-karpathy-skills — A single CLAUDE.md file encoding Karpathy's observations on LLM coding pitfalls. 52,803 stars. Proof that the most impactful "agent framework" might just be a well-written prompt contract. The viral growth (nearly 8K stars/day) says something about where developer attention is: not model weights, but behavioral scaffolding.
holaOS — Agent environment for long-horizon work, continuity, and self-evolution. 2,748 stars. Built for agents that need to maintain state across sessions and autonomously improve their own capabilities. Think of it as systemd for AI agents.
OpenMontage — First open-source agentic video production system. 12 pipelines, 52 tools, 500+ agent skills. 2,445 stars. If you're exploring AI video workflows beyond single-shot generation, this is the most complete open-source starting point. Directly competes with Runway's orchestration layer.
The Codex announcement is the one to watch closely. Not because screen monitoring is new — but because OpenAI is framing agents as persistent background processes that wake themselves up, schedule their own work, and run across days. That's not a copilot sitting in your IDE. That's a coworker with a cron job. Anthropic is winning on raw coding benchmarks with Opus 4.7, but OpenAI is betting the agent that lives on your machine wins the adoption game. Both are probably right about different segments — and that's what makes this the most competitive week in AI developer tooling since ChatGPT first hit the terminal.
— Aaron, from the terminal. Back next Friday.
Compare three approaches to AI agent browser automation. Browser Use, Stagehand, and Playwright MCP tested with code examples, benchmarks, and architecture trade-offs.
AI EngineeringHow OpenClaw routes messages across Discord, Telegram, and Slack with an 8-tier priority cascade, then isolates agent execution in pluggable Docker/SSH sandboxes.
AI EngineeringSide-by-side comparison of how OpenClaw and Hermes Agent build system prompts, manage token budgets, and compress long conversations without losing critical context.
AI Engineering