An agent treated its eval like a computer; today's tools just hand it one.

Firecracker microVM sandboxes for agents with no session clock — the maker's stated problem is that most sandbox providers kill your agent after 24 hours, which makes "autonomous for days" a lie at the infrastructure layer. Every agent and tool call runs in its own microVM; the filesystem is durable, so state persists exactly as the agent left it; sandboxes pause indefinitely and resume instantly; and snapshot, fork, and resume are first-class verbs — on our replay thread, where kassette forked workflow journals (07-14), the fork verb has now arrived at the machine layer. TypeScript and Python SDKs, an MCP integration, per-second billing ($0.0504 per vCPU-hour, $0.0162 per GiB of memory-hour), and a free tier with no card. Apache-2.0 and self-hostable, and the repo dates to last October — this is a launch of a built thing, not a promise. Reach for it when the agent's job outlives the day — migrations, long research runs, anything where "the sandbox expired" is the actual failure mode. Delete the cron job that re-provisions a fresh sandbox every morning and re-hydrates yesterday's state into it. Tradeoff: the Show HN landed at single-digit points, so community mileage is thin; per-second billing on an indefinite session is a budget you must cap yourself; and the same-day competition is real — see box in the footer, closed-source and claiming a tenth the price.

Hierarchical agent loops from a three-week-old org: a node iterates toward its goal in its own git worktree and spawns child nodes for separable subtasks, so the tree grows to fit the problem instead of a fixed plan. Every loop is bounded by hard caps — iterations, depth, children, cost, time — an operator can steer or stop any node live, and run metadata including spend lands in one local SQLite database with a terminal dashboard over it. `pip install plasma-fractal`, or install it as a skill through the Claude Code and Codex plugin marketplaces. Apache-2.0, 143 stars — a fresh bet by any measure, but with CI, docs builds, lint, tests, and codecov wired, which is more engineering hygiene than most month-old agent frameworks bother with. The lineage: 07-20's Shikigami put parallel agents in worktrees as a closed desktop beta; this is that shape arriving open-source. And Monday's Cursor swarm arithmetic — $1,339 to $10,565 for the same quality depending on model mix — is exactly why caps as first-class primitives matter more than orchestration cleverness. Reach for it when a task decomposes but you don't know into how many pieces up front. Delete the hand-rolled orchestrator script with its unbounded recursion and untracked spend. Tradeoff: the org is three weeks old and recursive self-organization is the feature and the risk in one sentence — the caps are load-bearing, so set them like you mean it.

An MIT TypeScript package (`libretto-browser-tools`) that gives any agent a browser harness: six tools, of which two do the work — `browser_snapshot` returns a compact accessibility tree instead of raw HTML, and `browser_exec` runs Playwright commands directly. The design bet is token economy: 55 percent lower cost than alternatives on complex pages, by the author's own benchmark. Browser backends are pluggable — local Chromium, Browserbase, Kernel, Steel, or Libretto's own cloud — with built-in adapters for Vercel's AI SDK and Pi. The lineage is the point: Libretto crossed our pool on 07-17 as the self-healing-Playwright startup in that day's verification cluster, and open-sourcing the harness underneath is the toolmaker giving away the moat's bottom layer while selling the reliability layer above it. Reach for it when your browser agent burns its context window on raw DOM — the compact-snapshot approach is the same bet Manifest made Monday, one layer down and self-hosted this time. Delete the Playwright MCP setup you've been apologizing for. Tradeoff: `browser_exec` executes model-written Playwright at runtime, which is power and attack surface in the same tool; adapters beyond AI SDK and Pi are "coming soon"; and the 55 percent figure is the vendor's own number until someone reproduces it.

A living wiki for your codebase, maintained by AI coding agents: plain markdown in the repo, indexed locally, reviewed in Git like any other change. The new part is the cadence machinery — setup installs three local launchd jobs: Sync scans your recent Codex and Claude Code conversations every five hours and queues useful knowledge into the relevant wiki, Garden reviews every wiki daily for stale, duplicated, or poorly connected pages, and Update keeps the CLI current. The stated scope is what the code can't say — decisions, invariants, gotchas, cross-service flows. Apache-2.0, YC S26, 582 stars since April; it first crossed our pool a week and a half ago and came back today as the day's strongest on-vertical Show HN (53 points). On the trail-as-asset thread this is the third beat: SkillOpt distilled rollouts into a skill (07-08), Grepathy extracted per-branch decision records (07-16), and this is the standing institution — scheduled distillation plus a gardener. Reach for it when the answer to "why is this shaped this way" lives in a transcript that expired two weeks ago. Delete the ARCHITECTURE.md nobody has updated since the person who wrote it left. Tradeoff: macOS-only, Codex or Claude Code only, Python 3.12+; a background job reading all your agent conversations is a scope grant worth thinking about (telemetry is opt-in, with an explicit never-sends list in the README); and a wiki is only as trustworthy as its gardener — Git review of wiki commits is the backstop, so actually review them.

A native macOS utility (Swift, about 2.5 MB, notarized) that puts a floating window on every Claude Code session and surfaces it the moment a session blocks on your input, then hides when all is clear. One click jumps to the exact terminal or editor pane — iTerm2, Terminal, VS Code, Cursor — and session data never leaves the machine. $4.99 a month or $48 a year after a seven-day trial; 84 votes and #15 on Product Hunt today. It is the smallest pick in the slate and the one that completes it: everything above moves the agent out of your terminal and lets it run longer, which quietly converts your job from typing to noticing. Our 07-06 watch asked for the unified away-from-desk inbox across machines; this is the at-desk version — six sessions, one signal. Reach for it when you run parallel Claude Code sessions and keep discovering one that has been waiting forty minutes for a y/n. Delete the cmd-tab patrol. Tradeoff: closed source, subscription-priced, Claude Code-only, macOS-only, and single-machine — the cross-machine, cross-harness inbox the watch actually wants is still unshipped, and at $48 a year, a free open-source clone is one motivated weekend away.

Off the thread but worth knowing: **box** (134 votes, #5 on Product Hunt) is the closed counterpart to Superserve, launched the same day — full Ubuntu VMs with admin rights, a web desktop, and SSH, provisioned in about two seconds by `box new`, snapshotting included, at $0.036 an hour, which the makers claim is roughly a tenth of E2B, Daytona, or Modal, on a no-dependencies stack they wrote themselves. **i-have-adhd** (MIT) was GitHub's #1 trending repo today at +1,866 stars — a skill that stops your coding agent from burying the answer, in ten rules ("lead with the next action," "cap lists at five," "no preamble, no recap, no closers"), installable from the plugin marketplace on Claude Code and Codex and loosely adapted from a clinical ADHD toolkit: the packaged-expertise thread (07-01) in its most human form yet, no diagnosis required. **Observal** (Apache-2.0, 2,223 stars since March) is a self-hosted control plane and system of record for your org's AI components — registry plus analytics, shipped via Docker and Helm — the fleet-scale sibling of today's single-agent slate. And Moonshot kept shipping past its own benchmark cycle: **kimi-cli** (10.6k stars, first crossed our pool Saturday, +329 today) and **kimi-code** (MIT, 4.5k stars) trended together on the same day Fireworks and Artificial Analysis published their Kimi-K3-versus-Fable numbers — model Friday, desktop Monday, CLI tooling today — while **Kastra**, our 07-10 runtime-authorization pick, launched on Product Hunt: the hook-policy-engine watch item starting its commercial arc.