The agent's hands finally left the browser.

Callstack — the React Native consultancy behind a sizable slice of that ecosystem's tooling — ships a device-automation CLI that lets coding agents open, inspect, and operate real apps on iOS, Android, TV, web, and desktop. The design statement is in the README: the CLI is the agent's hands, eyes, and evidence collector, not its brain. Instead of screenshot-guessing, agents get token-efficient accessibility snapshots with semantic refs like @e3, deterministic actions (tap, type, scroll, gesture, assert), and evidence captured on demand — screenshots, videos, logs, network traffic, crash context, React profiles. Sessions record to replayable .ad scripts with strict Maestro YAML export when a flow needs to live in CI. Works with native apps plus Expo, Flutter, and React Native; the authors' own framing is "Vercel's agent-browser, but for everything that isn't a browser." MIT, npm-installable, 3,410 stars, pushed today. Created end of January — this isn't a launch, it's the category's center of gravity crossing our pool.

Reach for it when the question is "did the agent's fix actually work on the device" — the agent verifies on a simulator or real hardware instead of reasoning about code and declaring victory. Delete the ritual where the agent says done and you tap through the app yourself to check. Tradeoff: it presupposes the whole simulator/emulator toolchain underneath, and it's deliberately hands-only — the judgment loop that reads the screen and decides what's next is still your harness's job.

An extension for the Pi coding agent that lets it operate macOS (14+) and Windows desktop apps: find windows, observe and search the visible UI, click, type, scroll, wait for changes. The mechanism is the pick: platform accessibility APIs and structured, state-scoped observations — eight tools like observe_ui and act_ui — rather than a vision model squinting at screenshots. The README's restraint is the most on-voice sentence of the day: "not a replacement for app APIs or MCP servers. If an app has a reliable direct integration, use that first. Computer use is most helpful when the only available interface is the app on screen." MIT, 1,388 stars, created in April, pushed today.

Reach for it when the workflow's only interface is a window — the license dialog, the settings pane, the vendor tool that never got a CLI — and it's the one manual break in an otherwise-automated loop. Delete the "I'll just click through this part myself" step. Tradeoff: it's an extension for one harness (Pi), so this is a reason to try Pi rather than a bolt-on for your current agent, and on macOS it wants Accessibility plus Screen Recording — most of the keys to the machine, the same grant we flagged on 07-02's macOS MCP pick.

Today's Launch HN (YC S26): managed Linux and Windows VMs an AI agent can see and drive — task runs that loop autonomously to a goal, workflows with branching and human-approval gates, low-level primitives (predict, ground, parse), and direct machine control down to SSH, VNC, and screenshots. The notable part is that the pricing is published like a utility's: $0.05 per step, machines at $0.05/hr Linux and $0.09/hr Windows, primitives at three to ten cents a call, free test keys, live today with no waitlist. Human takeover is built in — the agent pauses at CAPTCHAs and verification codes, a person clears the blocker, the run resumes. SDKs for Python, Node, Go, Ruby, and PHP.

Reach for it when the workflow lives in software you can't put hands into — the vendor portal, the Win32 ERP, the claims system from 2009 — and you'd rather rent the machine than babysit the browser. Price the back-office data-entry contract against a nickel a step and see which survives. Tradeoff: this is exactly the screenshot-loop school the two picks above engineered away from — hosted, closed, metered per step, and the meter runs while the agent is confused. Your legacy system's credentials also now live in someone else's VM; the approval gates exist for a reason.

Email infrastructure built for agents as the customer: marketing and transactional sending, one-to-one conversations with human escalation, agent inboxes on custom domains (beta, by request), all headless via API and MCP — there is deliberately no dashboard. The detail that stops you mid-scroll: point any agent at nitrosend.com/SKILL.md and it signs itself up, onboards, connects your domain, and sends. The maker is ex-SmartrMail with billions of messages behind him, so the unglamorous deliverability plumbing is his home turf, not a weekend learning exercise. Launched today, #5 on Product Hunt; three months of Pro free at launch.

Reach for it when your agent's job currently ends where email begins — the receipt, the notification, the reply that needs sending at 3 a.m. without a human relaying it through an ESP built for marketing teams. Delete the glue code between your agent and an email platform that assumes a person is logged in. Tradeoff: an agent with an outbound mail server is the most efficient sender-reputation-destruction machine ever built, and the deliverability sins land on your domain. And self-signup-via-SKILL.md is a great demo that quietly removes the human from a vendor decision — the inbox half is also still beta-by-request, so the full loop isn't general yet.

The picks above put agent hands on human interfaces; Cito is the other answer — for when a machine interface exists but was rate-limited for human politeness. It's academic search over 236M papers (146M with SPECTER2 dense vectors), hybrid keyword-plus-vector retrieval fused with RRF and reranked by a cross-encoder, returning abstracts, citation counts, DOIs, and open-access PDF links. Free web search, a free JSON API at 100 requests/minute, and a native MCP endpoint for Claude Code and friends. The origin story is one sentence: "built because every academic API throttled my agents to death" — Semantic Scholar's official limit is one request per second, which stalls any real deep-research pass.

Reach for it when your research agent spends more wall-clock sleeping between requests than reading. Delete the sleep(1) in your literature pipeline. Tradeoff: one person's free side project is now load-bearing infrastructure for your agent — no SLA, upstream corpus licensing unaddressed, and rate limits this generous last exactly as long as they aren't abused. Fine for the research loop; don't build the product on it yet.

The counterweight the rest of today requires. When agents act everywhere, the what gets logged; the why lives in a session transcript Claude Code deletes after thirty days by default. Grepathy (npx grepathy init) reads your transcripts locally, extracts the decisions the agent made along the way, and writes them to .ai/why/<branch>.md, committed with the code — each entry tagged with status, including the damning one: "agent-initiated — not requested in plan or prompts." It runs itself off hooks (fires on push, never blocks, never touches staging), and future agents meet the history automatically — a hook injects the relevant entries right before an agent edits a file that has one. The origin story is the entire pitch: the author's agent silently pre-created guest users in Clerk on a contract job, the CTO asked why, and nobody — including the author — knew.

Reach for it when the review question is "why was this done this way" and the honest answer is "the agent decided and I didn't notice." grep -rn "agent-initiated" .ai/why/ returning a list of every decision nobody approved is the feature. Tradeoff: nine days old, 47 stars, and the value rides entirely on extraction quality — a summary of a transcript is testimony, not evidence, so it's a reading aid for the record, not a replacement for keeping it.

Off the thread but worth knowing: grok-build — xAI open-sourced its coding agent harness (Rust, fullscreen mouse-interactive TUI, Apache-2.0) and it pulled 9,185 stars in two days; it's the Nth open coding CLI and notable mainly for who shipped it, but it's installable and you can judge the harness instead of sniffing its packets, which is more than could be said on 07-13. StyleSeed (Show HN today, 764 stars, MIT) packages 74 design rules, brand-flavored token skins, and a scored quality gate so agent-built UI stops looking generated — the ui-skills instinct productized, and fittingly, ui-skills itself (our 06-09 edition) is back on trending today. Also re-trending: anthropics/knowledge-work-plugins (22.8k stars, created in January) — Anthropic's plugin library for Claude Cowork, riding the same what-you-load-in wave as the 06-19 skills edition. And one for the systems crowd: misa77 (99 stars, MIT, C++) claims 1.5–3x faster decompression than LZ4 at better ratios, with the honest caveat that compression is slow.