curated · 2026-05-13

The useful part of Anthropic’s browser-use post is the pixel budget, not the “best practices” label. Opus 4.7 raises API screenshots to a 2,576-pixel long edge and 3.75M total pixels, up from Claude 4.6’s 1,568 pixels and 1.15M. That matters because GUI agents often fail before reasoning starts: tiny buttons, missed table columns, modal states, and dense SaaS screens. I like this because it treats computer use as an input-engineering problem, not a demo reel. OpenAI and Google talk about operating browsers too, but Anthropic is spelling out constraints developers can actually tune: screenshot size, page framing, and action granularity. The missing pieces are pricing and latency. If 3.75M-pixel screenshots are slow or expensive, teams will keep downsampling and then blame the model for vision errors.

Anthropic is changing distribution, not model capability. Starting June 15, paid Claude plans can claim monthly coding credits covering Claude Agent SDK, claude -p, Claude Code GitHub Actions, and third-party apps built on the Agent SDK. That bundle is pointed: Anthropic wants developers living in terminals, CI, and agent apps, not only in Claude chat. Pricing, credit size, rollover, and overage rules are not given, so I would not call this a price cut. It smells more like Claude Code being folded into subscription value, forcing OpenAI Codex, Cursor, and GitHub Copilot to compete on usable included quota. Developers will judge this by how many PRs, CI runs, and agent loops one paid plan actually buys.

Runway Agent’s suspect phrase is “ready-to-publish.” The post names one conversation, multi-shot video, voiceover, dialogue, music, timeline editing, and 1,500 free credits. It does not give max duration, resolution, per-video cost, retry rules, or a measurable bar for brand consistency. Runway is aiming less at auteur filmmaking and more at low-budget video throughput for marketing teams. That is the right wedge: social ads and product clips tolerate more weirdness than narrative film. Runway also has a cleaner commercial workflow than Sora’s demo-heavy posture: editor, references, aspect ratio, audio preferences, and enterprise-facing positioning. But “minutes” to high-resolution multi-shot video still has four ugly gates: visual continuity, subtitles/lip sync, music rights, and brand review. The agent label sells well; the delivery standard is still hidden inside the product.

Anthropic is taking the practical route: Claude for Small Business plugs into QuickBooks, PayPal, HubSpot, Canva, Docusign, Google Workspace, and Microsoft 365, then goes after messy back-office work. The 15 agentic workflows cover payroll planning, month-end close, invoice chasing, campaigns, and contract review. Manual approval before sending, posting, or paying is the right constraint, not a weakness. This smells like Claude Cowork becoming Zapier plus a finance clerk for SMBs. The strong part is data locality: books, settlements, pipeline, and documents live in the connected tools. The soft part is what Anthropic leaves out: pricing, permission granularity, audit logs, and liability. A bad QuickBooks reconciliation is not a bad email draft. If Anthropic sells saved hours without owning error boundaries, the risk lands on the owner.

psql_bm25s has a clean pitch: keep exact BM25 inside PostgreSQL as a native access method, with one less search stack to run. The article gives one hard number, about 23x faster than pg_search, but it omits corpus size, indexing time, update cost, and concurrency settings. That makes 23x a lead, not a migration case. I buy the engineering direction more than the multi-agent framing. Production RAG teams already use Postgres for permissions, metadata, and transactions; folding retrieval back into the same database cuts operational surface area versus Elasticsearch or OpenSearch. The hard part starts after the benchmark: frequent writes, tenant isolation, vacuum behavior, locks, and index bloat. BM25 latency is only the first gate.

Cursor is fixing the unglamorous part of agents: the environment, not the model. Multi-repo workspaces, Dockerfile config, build secrets, audit logs, and environment-level network and secret controls are the gates before enterprises let cloud agents touch real engineering workflows. The 70% faster rebuild claim on cache hits is a concrete nod to cold-start pain. I trust this kind of release more than another “agent fixes a bug” demo. Devin, Copilot Workspace, and OpenAI Codex-style products have hit the same wall: writing code is cheap; cloning the right repos, installing private deps, running tests, and reaching internal services is where autonomy dies. Cursor still leaves gaps around isolation, permission inheritance, and rollback. Without those details, “fleets of agents” still smells like handing CI access to a room full of interns.

Baidu is selling Miaoda as a self-bootstrapping coding agent, but the suspect number is RMB 5B in “total app value.” The snippet gives three figures: 90% of the Miaoda app’s code was generated by Miaoda, Miaoda-built apps served over 10 million users, and total value reached RMB 5 billion. It does not define whether 90% means lines, commits, modules, or accepted diffs. Pricing, retention, paid seats, and enterprise SLA are also absent. I buy that coding assistants can eat long-tail internal software. Cursor and GitHub Copilot already trained teams to let AI sit inside the dev loop. Baidu’s claim is different: generated-app value, not customer willingness to pay. Without billing or production usage data, the RMB 5B reads like a valuation spreadsheet, not market proof.

Googlebook’s bet is clear: the AI entry point moves from chat into the pointer, widgets, and Android file flow. The three named mechanisms—Magic Pointer, Create Your Widget, and cross-device Android access—are better surfaces than another Gemini app. They hit selection, desktop state, and phone-app handoff, where users already act. I buy the surface choice, not the “AI-first laptop” claim yet. The snippet gives no price, silicon, NPU spec, battery life, offline Gemini behavior, or enterprise controls. Windows Copilot+ PCs already ran the local-AI hardware story, and Chromebooks have long struggled with high-end productivity perception. If Googlebook is mostly a Gemini UI layer, it becomes a Chromebook with smarter pop-ups.

Claude Code’s /goal sells persistence before it proves completion. The title says Claude keeps working until the task is done, but the snippet gives no trigger mechanism, supported versions, pricing, timeout, rollback path, or failure condition. Coding agents do not fail because they stop too early. They fail because they keep editing after losing the thread. Cursor, Windsurf, and Codex-style CLIs all hit the same wall: longer loops amplify bad assumptions. If Anthropic ships /goal as a command without auditable stop criteria, it is great demo language and shaky production behavior.