~/reading

42

Issue 42/October 17, 2026/6 tools · 2 headlines

MicroVM sandboxes and open control planes that let power users own the agent loop

Today's picks center on hardware-isolated runtimes, programmable egress, and cloneable harnesses that sit under any model. Power users get disposable kernels, policy outside the agent, and eval loops they can wire themselves instead of waiting on closed IDEs.

01 — Power tools

The stack

  1. 01repoGitHub
    SPACE↗

    Perplexity AI

    Firecracker microVM sandbox platform with HTTP API, pause/resume, snapshots, and host-side credential and egress control.

    Use caseUse it to give every agent task its own kernel-isolated Linux environment that boots in tens of milliseconds and never sees real secrets.

    sandboxmicrovmfirecrackerisolation
  2. 02harnessGitHub
    Sandcastle↗

    mattpocock

    TypeScript library that orchestrates coding agents inside isolated sandboxes with configurable branch strategies and merge-back.

    Use caseUse it to parallelize multiple AFK agents across Docker/Podman/Vercel sandboxes and automatically fold their commits.

    harnessorchestrationtypescriptsandbox
  3. 03toolGitHub
    Stately Agent↗

    statelyai

    XState-powered agents where the state machine owns control flow and the model only proposes legal events.

    Use caseUse it to make invalid agent actions impossible by encoding the entire loop as an inspectable, resumable state machine.

    harnessstate-machinexstatedeterministic
  4. 04repoGitHub
    A-Mem↗

    WujiangXu

    Agentic memory system that dynamically organizes notes with Zettelkasten-style linking and continuous evolution.

    Use caseUse it to let an agent itself decide how to index, link and refine long-term memories across sessions.

    memoryagenticknowledge-graph
  5. 05repoGitHub
    Meta-Agent↗

    canvas-org

    Continual harness optimizer that turns execution traces into prompt, tool and control-flow improvements without weight updates.

    Use caseUse it to automatically evolve the scaffolding around a frozen model by evaluating candidate harness changes.

    harnessmetaoptimization
  6. 06repoGitHub
    Agent Diff↗

    agent-diff-bench

    Interactive sandboxes that replica third-party APIs (Slack, Linear, etc.) for deterministic agent evals and RL.

    Use caseUse it to train or score agents against realistic API state changes without side effects or rate limits.

    evalsandboxrlapi

02 — In the wild

Articles and releases

  1. releaseMarkTechPost
    Perplexity open-sources SPACE, its production Firecracker agent sandbox↗

    Perplexity / MarkTechPost

    SPACE, the sandbox behind every Perplexity Computer session, is now Apache-2.0 on GitHub. It gives each task a Firecracker microVM with host-controlled credentials and egress, median 60 ms spin-up, pause/resume and an E2B compatibility wrapper.

    releasesandboxmicrovmperplexity

    Shipped

  2. articleTech Thought Leaders
    The agent sandbox has moved into the operating system↗

    Ivan Tarin

    Microsoft Execution Containers GA turns agent containment into a Windows platform feature with process, session and MicroVM backends. Policy now lives outside every framework, so enterprises buy isolation once instead of per-agent.

    articlesandboxoswindows

    Read

Get the daily tools digest

RSS