The platform vendors just took over writing your agent's onboarding.

Google's Android team publishes a first-party skills repository: AI-optimized, modular instructions following the open agentskills.io standard, installed with android skills add --all into every coding agent the CLI detects on your machine. The scoping decision is the load-bearing detail — they write skills only for the areas where their own evals show LLMs underperform (an r8-analyzer skill, not another Compose tutorial), which is the first vendor skills repo we've seen that starts from measured model failure rather than marketing surface area. Apache-2.0, 6,236 stars, created in March and trending today. Lineage worth noting: google/skills crossed our pool on 06-08 and again Monday; now the platform teams inside Google are shipping their own. That's not one vendor experimenting, that's an org habit forming.

Reach for it when your agent writes Android and gets wrong exactly the parts that don't fit in training data — build tooling, optimization, the current official line. Delete the pinned "how we do Android here" prompt you cargo-culted from a gist. Tradeoff: "Public contributions are not accepted" — this is a one-way channel carrying developer.android.com's official position, so where best practice is contested, your agent now confidently holds Google's side of the argument.

AWS's official, GA toolkit for coding agents: MCP servers, skills, and plugins covering service selection, CDK/CloudFormation, serverless, containers, billing (aws-core), building agents on Bedrock and AgentCore (aws-agents), analytics, and a devsecops pack that investigates incidents and scans code. The distribution is the story more than the contents: /plugin install aws-core@claude-plugins-official — AWS publishes its onboarding through Anthropic's default plugin marketplace, plus Codex's marketplace, Cursor team marketplaces, Kiro, and aws configure agent-toolkit in the AWS CLI itself. The era of documenting for humans and hoping the agent read it is over; the vendor now packages for agents and publishes where agents shop. Apache-2.0, 1,925 stars, created in April; second crossing of our pool (first on 06-25).

Reach for it when your agent's AWS work runs on half-remembered CDK patterns and a community MCP server of unknown provenance. Delete that community server. Tradeoff: this is AWS teaching your agent to build on AWS — the "service selection" skill is authored by the seller of the services, so the guardrails and the lock-in ship in the same box.

LM Studio — the desktop app a large share of the local-model crowd already runs — ships Bionic, a separate agent application for open models: coding against local codebases, research with native web search, document and presentation work, sandboxed file access, and voice input via local Voxtral transcription. Local models you've downloaded run on your hardware; a hosted menu of open frontier models (GLM 5.2, Kimi K2.7 Code) bills through their Secure Cloud with a stated zero-data-retention commitment. This is the 06-30 watch, shipped: we kept cutting local-model launches and said the pick would be the installable tool riding the wave — a local-first agent harness, not another checkpoint. Here it is, from the vendor with the distribution to make it matter. 264 points on the HN front page.

Reach for it when you want agent work — code, documents, research — without an API bill or your files leaving the machine. Delete the chat-app-plus-copy-paste ritual that currently passes for local-model workflow. Tradeoff: the app is closed source, and the convenience curve bends toward the metered cloud tier — the truly-local story is bounded by your VRAM, and the day's loudest model (Kimi K3) is exactly the kind of thing that menu exists to sell you.

A simulated bash environment with an in-memory virtual filesystem, written in TypeScript, designed for AI agents: npm install just-bash, instantiate new Bash(), and your agent gets grep, sed, awk, find, jq, sqlite3, yq — plus opt-in curl, Python, and QuickJS — against a filesystem that isn't your machine. Custom TypeScript commands compose with pipes and redirection, so your app's own operations become unix-style verbs the agent already knows how to chain. The premise is quietly sharp: the agent's most fluent interface is a shell, and the safest shell is one with nothing real behind it. Version 3.1.0, 3,946 stars, created in December, a THREAT_MODEL.md in the repo root, Apache-2.0 declared in the npm package (no LICENSE file at the repo root — worth a glance if your org checks that box).

Reach for it when the agent needs to slice files and data mid-loop — in serverless, in the browser, in CI — and a real shell is impossible or inadvisable. Delete the homegrown read_file/write_file/search tool suite you built to avoid exactly this. Tradeoff: simulated bash is not bash, and the edge cases where they differ are precisely the ones an agent will find; it's beta by its own label, and every opt-in runtime widens the surface that threat model exists to catalog.

The counterweight today's slate requires. Four picks of vendors loading capability onto your agent — skills from Google, plugins from AWS, a shell, an app — and Ratel is the layer that keeps it all out of the context window until needed: it indexes your tools and skills into a catalog the agent queries per turn, injecting only the matching capabilities instead of front-loading every schema. In-process BM25 plus semantic retrieval, no vector database, SDKs on npm, PyPI, and crates.io, published benchmarks claiming roughly 80% token reduction. Show HN today from its two authors; created in November, 209 stars, Apache-2.0 and MIT. The demand side is the 07-13 teardown that found a stock harness spends some 24,000 tokens on tool schemas before reading your prompt — and today the supply side got materially bigger.

Reach for it when the agent's tool list has outgrown its judgment — wrong-tool calls climbing in lockstep with your token bill every time someone adds an MCP server. Delete the deferred-tool-loading glue you were about to write yourself. Tradeoff: a retrieval layer between the agent and its capabilities is a new failure mode — the tool it doesn't surface is a tool the agent doesn't have; and the benchmark numbers are the authors' own, not independently reproduced.

Off the thread but worth knowing: anthropics/cwc-workshops (1,496 stars, Apache-2.0) trended today — Anthropic's Code with Claude workshop materials, nine hands-on builds from eval-driven agent development to memory ("goldfish to colleague" in 45 minutes); explicitly unmaintained and not a product, but the best free agent curriculum in the pool, and the same first-party pattern as the slate: the lab ships its pedagogy. Agent Substrate (776 stars, Apache-2.0, two months old, "not an officially supported Google product") multiplexes ~250 stateful agent sessions onto 8 Kubernetes pods with suspend/resume and gVisor isolation — one for the fleet-ops shelf from 06-30. Forall (276 stars, Apache-2.0) is a coding agent that generates machine-checkable proofs alongside TypeScript, Java, and Rust — CLI or a verify-only MCP mode for Claude Code and Cursor, with the caveat that verification runs hosted behind an API key; the 07-15 Lean-verified fleet suggested this was coming to working code. And code-review-graph re-trended (19,600 stars — first crossed our pool 05-19): a local code-intelligence graph for MCP and CLI, working the same context-reduction lane as Ratel from the codebase side.