~/reading

19

Issue 19/September 21, 2026/9 tools · 3 headlines

Sandbox-first harnesses that fence the agent, not the IDE

Today’s clone-and-run stack is isolation and control planes under the model: WASM cells, OS seatbelts, KVM/K8s sandboxes, and Rust fleet orchestrators. Power users wire these into loops instead of betting on another chat shell.

01 — Power tools

The stack

  1. 01repoGitHub
    Boxed↗

    akshayaggarwal99

    Self-hosted sovereign code-execution substrate with Go control plane, Rust in-sandbox agent, BYOK auth, and Docker/gVisor-hardened ephemeral sessions.

    Use caseUse it to spawn fail-closed Docker sandboxes from Python or TypeScript SDKs and stream artifacts over JSON-RPC.

    gorustdockercontrol-plane
  2. 02harnessGitHub
    POCKET Agent↗

    ItsNotAILABS

    Long-running repository agent with durable goals, recursive agent harnesses, WASM capsules, budgets, leases, receipts, and detach/reattach sessions.

    Use caseUse it to set a multi-hour repo goal, run autonomous work with budgets, detach, and reattach with verifiable receipts.

    long-runningwasmreceiptscli
  3. 03toolGitHub
    Anthropic Sandbox Runtime (srt)↗

    anthropics

    OS-level sandbox CLI and library using Seatbelt, bubblewrap, and Windows filters plus proxy-based network allowlists without containers.

    Use caseUse it to wrap MCP servers or agent bash with filesystem deny/allow rules and domain-scoped egress via srt.

    os-sandboxseatbeltbubblewrapmcp
  4. 04toolGitHub
    sandbox-runtime-rs↗

    Ghostlock-AI

    Rust port of Anthropic sandbox-runtime aiming for much faster startup and smaller binaries with Seatbelt profiles and HTTP proxy filtering.

    Use caseUse it to drop a single native srt binary in front of agent processes when Node overhead is unacceptable on macOS.

    rustseatbeltperformanceport
  5. 05repoGitHub
    meclaw↗

    mmeyerlein

    Ontology-grounded agentic build system in one Rust binary: actor cells with SQLite, sandboxes, vault, voice, and web surfaces for composing agentic OS graphs.

    Use caseUse it to grow a message-routed multi-cell agent system where each actor owns its sandbox and secrets stay in a vault cell.

    rustactoragentic-ossandbox
  6. 06harnessGitHub
    EnvHarness↗

    google-research

    Programmable layer that wraps frozen training environments with plugins so EnvRigger can reshape tasks toward agent weaknesses without changing verifiers.

    Use caseUse it to co-evolve SWE-bench or WebArena-style envs with your policy via reset/step bridges and targeted component banks.

    trainingevalrlgoogle
  7. 07toolGitHub
    landstrip↗

    landstrip

    OS-level sandbox for coding agents using Landlock, Seatbelt, and AppContainer with JSON policies and drop-in CLI for command runs.

    Use caseUse it to run cargo test or agent tools under a parametrized filesystem and network policy without a full container.

    landlockpolicyclicross-platform
  8. 08harnessGitHub
    defending-code-reference-harness↗

    anthropics

    Reference vuln-finding pipeline that forces every agent into gVisor with egress allowlist proxy and never mounts credentials into agent reach.

    Use caseUse it to study or fork a production-shaped sandboxed agent loop for crash-finding with bin/vp-sandboxed entrypoints.

    gvisorsecurityreferencepipeline
  9. 09repoGitHub
    AHE (Agentic Harness Engineering)↗

    china-qijizhifeng

    Observability-driven automatic evolution of coding-agent harnesses with evaluate-analyze-improve loops and reported Terminal-Bench lifts.

    Use caseUse it to evolve harness components over iterations inside E2B sandboxes and freeze a transferable harness.

    evolutionevale2bresearch-origin

02 — In the wild

Articles and releases

  1. releaseVentureBeat
    Google open-sources EnvHarness for adaptive agent training worlds↗

    VentureBeat

    Google Cloud AI Research released Apache-2.0 EnvHarness code that wraps static benchmarks with plugins so environments adapt to agent weaknesses. Reported gains reach about 9 points on held-out tasks with fewer steps across SWE-bench Verified and related suites.

    envharnessgoogletrainingopen-source

    Shipped

  2. articleForkast
    Google ships managed agent harness credentials that never enter the sandbox↗

    Blair Hayes

    Google’s Managed Agents preview routes auth through an egress proxy that injects headers so secrets never sit inside the agent Linux sandbox. The piece maps this against other September credential models from Rubrik, Microsoft, and NVIDIA.

    credentialssandboxgooglesecurity

    Read

  3. releaseGitHub Releases
    Anthropic Sandbox Runtime v0.0.77 tightens root and deny rules↗

    anthropics

    srt v0.0.77 (mid-September 2026) adjusts Linux write-root and macOS read-root behavior so broader allows no longer clear nested denies. The OS-level seatbelt and bubblewrap tool remains the lightweight fence many agent loops wrap around MCP and bash.

    srtanthropicreleasesandbox

    Shipped

Get the daily tools digest

RSS