Grok Bot vs Muse: What Happens After Prompt Injection?
A look inside Grok Bot and Muse's cloud computers: how command review, network checks, and credential isolation limit what happens after prompt injection.
How We Built Errand with One Developer in One Week
I built Errand, an open-source alternative to Grok Bot and Muse, by myself in about a week. Confession: I didn't build the hard parts. Here's what Runta already had, what I wrote, and where the week actually went.
Introducing Errand: an open-source alternative to Grok Bot and Muse
Errand gives an AI agent its own computer to run tasks while you’re away—open-source alternative to Grok Bot and Muse.
Introducing Exo Harness for Runta Runtimes
Exo Harness on a persistent Runta runtime, with protected model credentials and a workspace you can pause and resume.
Introducing FrontierHarness Eval
FrontierHarness eval is a harness benchmark of 9 production coding harnesses on 30 agentic tasks—same model and runtime. Pass rates cluster; cost spans 17x.
Introducing SSH access for Runta runtimes
Connect any SSH-capable agent to a persistent Runta runtime, with access to services but without exposing the underlying credentials.
Agents aren't software. A conversation with Jeff Dean helped me see why.
Reliability was designed at write time because code was deterministic. Agents break that assumption, and the layer beneath them, from checkpointing to recovery, has to be rebuilt around a new question: what does it mean to resume a half-finished judgment?
Runta: the execution layer for agents
Runta is the execution layer for AI agents: computers they run on, secrets they can use without seeing, and access you govern. $20M seed led by Martin Casado at Andreessen Horowitz.
Introducing ClawShell, The Security Layer OpenClaw (Peter) Needs
ClawShell enforces OS-level isolation for AI agents — so even if the agent is fully compromised, it can't access what it was never given.