One of our engineers built Errand, our open-source take on Grok Bot and Muse, and wrote up how it came together. One developer, one week. The app is ~4k lines of TS. Every agent still gets its own machine with a shell, a browser, and a filesystem that survives between tasks. He never wrote a provisioner or an event log infra. If you're about to build this layer yourself, that's what Runta is. Week one goes into your product instead of a provisioner. I'd like to hear where you're stuck, if you're in the middle of it right now.

Grok Bot and Muse are great. We built an open-source version and called it Errand. Hand it a task, close the laptop, come back in the morning.

Errand took one developer about a week to build. An agent product like this usually means building a pile of infrastructure first: persistent machines, file storage, tracing, credential handling, checkpoints, token budgets. None of it had to be built for Errand. Runta already had it, and Errand just runs on top. Not having to build the infrastructure into the app means you get to spend your time on the parts end users actually touch: integrations, workflows, and how the whole thing feels to use.

Exo Harness now runs natively on Runta. In FrontierHarness Eval, Exo was cheapest per completed task at $1.05. On the hardest task it hit its 51-step cap and quit at $1.46 while others kept spending. That is the harness you let rewrite itself. Exo's Executor holds no durable state, so the agent can modify it. History, artifacts, secrets and sandbox lifecycle sit in the Harness, out of reach. Runta provides a resumable environment, so the Exo Harness can self-evolve freely. Exo never holds the model API key, only a stub. Our egress gateway injects the real one at the provider. Code the agent wrote can read whatever the Executor can, and all the Executor has is a stub.

Wondering where your harness would land on the FrontierHarness leaderboard? Last week the most common reply was "can I run mine on this." Now you can. Bring your harness and your own model key, and it runs on Runta under the exact conditions the leaderboard was built on: same gateway, every run a fresh restore of the same golden checkpoint, memory state included. See where your harness actually lands next to the rest of the field. Send us the run, and if it stands out we'll include it in the next official round.

Want to try it? Reply with your harness repo and DM me your Runta account email. You bring the model key, we cover the runtime: I'll load some free Runta credits to get you started. Or you can directly request demo via: runta.com/request-demo

A year ago the question was which model. Now it's which harness. Pi, Exo, Claude Code, Codex, DeepSeek Harness and 4 others. Same model, same tasks, same runtime. 360 runs, 2 billion tokens. Pass rates: 50% to 67%. Cost per pass: $1.05 to $18.34. Introducing FrontierHarness Eval. 🧵

Something we learned: With implicit prefix caching, running a task once during debugging leaves it warm for hours. Test a harness on Tuesday, benchmark it Wednesday, and it shows up cheaper than it should. Nothing in the logs tells you why. So no benchmark task was touched before the formal run, and every one of the 360 trials is a fresh restore of one checkpoint, memory state included.

Here is the run that made us write this up. One of the hardest tasks, the pass rate (all model efforts) on DeepSWE is 38%, Pi and Claude Code both fixed it. Pi took 90 turns and $2.50. Claude Code took 381 turns and $64.36. Roughly 26x more for the same fix.