Runta (@runta) header
Runta (@runta) profile picture

The execution layer to control what AI agents can do as they run.

San Francisco Bay Area
Joined September 2025
Photo by Runta (@runta)Photo by Runta (@runta)Photo by Runta (@runta)Photo by Runta (@runta)Photo by Runta (@runta)Photo by Runta (@runta)Photo by Runta (@runta)Photo by Runta (@runta)
Pinned Tweet
If you're trying to give agents real work and control is what's stopping you, apply for early access today. Every new account starts with $50 in free credits, enough to run an agent for a month, with no credit card needed.
Today we're announcing @runta's $20M seed, led by @a16z. Software is constrained when you write it. Agents have to be constrained while they run. Runta is the execution layer that controls what AI agents can actually do.
7
2
18
7,329
Runta retweeted
Is there a best harness for each popular open model? We partnered with @runta to build FrontierHarness Eval v1.1, built by running 83B tokens across 22 harnesses and 5 open models. We found that cost per pass can vary by up to 4.4x across harnesses. If you’re running a coding agent with open models, maintain your own evals and re-test your harness on new model releases.
FrontierHarness v1.1 results: when you switch models, consider switching harnesses too, at least if you care about your token bill. The full run: 83B tokens across 27,280 agent runs, each in its own @runta runtime, inference by @baseten. Codex CLI ranks 1st on cost with GLM-5.3, and 10th of 22 with Kimi-K3. On any one model, cost per pass varies by up to 4.4x across harnesses. Only Pi never left the top 6 on cost. 🧵
16
4
1
57
4,614
great chart from @runta
1
1
42
On the same model, agent harnesses differ by up to 4.4x in cost per pass. And 64% of inference spend went to runs that failed. FrontierHarness v1.1, the largest harness × model eval we know of.
FrontierHarness v1.1 results: when you switch models, consider switching harnesses too, at least if you care about your token bill. The full run: 83B tokens across 27,280 agent runs, each in its own @runta runtime, inference by @baseten. Codex CLI ranks 1st on cost with GLM-5.3, and 10th of 22 with Kimi-K3. On any one model, cost per pass varies by up to 4.4x across harnesses. Only Pi never left the top 6 on cost. 🧵
1
5
243
Cursor/Notion recent incidents showed how stored agent traces can leak credentials. Agent traces and logs can leak. If an agent can read a API key, that key can leak too. At Runta, credentials never enter the runtime. The VM only gets a stub. All outbound traffic goes through our egress gateway, which adds the real key only for the hosts and paths you allow.
3
1
8
557
There’s a surprise 22nd harness in the FrontierHarness v1.1 eval, running alongside the other 21 under the same conditions. We’ll reveal it with the final results. Hint: it’s racing a version of itself. Any guesses?
10
2
20
2,519
Halfway through FrontierHarness v1.1, the leanest harness costs the least per pass on most models. Pi uses the fewest tokens per trial on all 5 models, 550k to 900k, while others run up to 6M. It ranks #1 on cost per pass on 4 of the 5. Pi ships 4 tools (read, bash, edit, write) and a ~650-token system prompt. Pi + DeepSeek-V4.1-Flash has the highest pass rate so far, close to 80%, at $0.06 a pass. Overall, Codex still leads on cost per pass, carried by Kimi K3. Is lean winning because it’s better designed, or because today’s models no longer need thick scaffolding?
9
3
42
6,909
FrontierHarness v1.1 is past halfway. 21 harnesses, 5 models, 62 tasks. Cost per pass: 1. Codex $0.57 2. Pi $0.67 … 21. Antigravity $1.32 First vs last: 2.3× cost per pass, 1.9× average cost per task. Pass rates: 56.2% to 77.6%. Which gap would actually make you switch harnesses, pass rate or cost?
5
1
11
1,390
We ♥️ Harbor
awesome work, built & run with harbor harbor x runta
1
7
369
Live board: frontierharness.org/live/ Every run executes in a @Runta runtime under the same limits, and shows up with its outcome, attempts, and token cost as soon as it finishes. Rankings will shift as runs come in. Each harness × model pair is final after 248 runs. The 62 tasks span software engineering (20), science and math (10), systems (8), data (7), ML (6), security (5) and 6 in other domains. Harnesses include Claude Code, Codex, Pi, Oh My Pi, Hermes Agent, Cline, ZCode, KimiCode, DeepSeek Harness, Grok Build, Antigravity, OpenCode and Copilot.
7
13
6
116
53,029
Runta retweeted
best GLM-5.3 harness -> KimiCode best Deepseek-v4.1-flash harness -> ZCode best Kimi-K3 harness -> Codex ...so open-weight models work best on "other" vendor harness...
We’re spending up to 100B tokens across 26,040 agent runs to answer one question: when you switch models, should you switch harnesses too? Our v1.0 FrontierHarness leaderboard ranked harnesses on a single model. V1.1 expands it to 21 harnesses × 5 models × 62 tasks × 4 attempts, 72x the size of v1.0. Results are coming in live. Link below🧵
17
4
62
5,702
Runta retweeted
What’s the best harness for an open model on coding and agentic tasks? We partnered with @Runta to investigate: 100B tokens and 26k trials later, the final results are coming in. Full report next week.
We’re spending up to 100B tokens across 26,040 agent runs to answer one question: when you switch models, should you switch harnesses too? Our v1.0 FrontierHarness leaderboard ranked harnesses on a single model. V1.1 expands it to 21 harnesses × 5 models × 62 tasks × 4 attempts, 72x the size of v1.0. Results are coming in live. Link below🧵
2
5
44
3,788
Runs are live. Which harness will finish FrontierHarness v1.1 with the highest pass rate, and which with the lowest cost per pass? Reply with your two picks. We’ll draw 5 winners from correct entries for $100 in Runta credits each. Entries close Nov 12, 9:00am PT.
We’re spending up to 100B tokens across 26,040 agent runs to answer one question: when you switch models, should you switch harnesses too? Our v1.0 FrontierHarness leaderboard ranked harnesses on a single model. V1.1 expands it to 21 harnesses × 5 models × 62 tasks × 4 attempts, 72x the size of v1.0. Results are coming in live. Link below🧵
4
1
4
663
Runta Singapore Grand Opening!
Replying to @runta
@Runta's Singapore office is open! 🇸🇬 Six months of preparation. Skipped the cramped high-rises for a green, light-filled workspace. Hiring for agent infra team for both SG and US: * Infra Engineer (FaaS / container scheduling / virtualization / RL infra) * Product Engineer (build in public, open-source, community, design sense) DM me if you’re into agent infrastructure.
1
7
567
Runta 10/10/2026 is now available! Resize on cold checkpoint restore. Restore a cold checkpoint with a different CPU and memory shape, so you can scale the runtime without rebuilding its environment. Smoother runtime creation at scale. Improved batching, scheduling, and higher API limits help runtime creation stay responsive during bursts. Simpler ChatGPT subscription setup. New runtimes use your account’s default model, while unavailable subscriptions stay hidden.
1
1
6
649
Pi 1.0 harness support is ready at Runta. Code-mode + native MCP is finally here:
Pi 1.0 runs on Runta today. Each agent gets its own machine: hands-on in a browser terminal, or headless over an API with live streaming, steering, and cancel. Pi never sees your real keys. Our egress layer swaps in model and credentials on the way out.
1
6
618
Runta retweeted
Pi 1.0 runs on Runta today. Each agent gets its own machine: hands-on in a browser terminal, or headless over an API with live streaming, steering, and cancel. Pi never sees your real keys. Our egress layer swaps in model and credentials on the way out.
4
2
2
32
5,128
Runta retweeted
Your agent doesn't need to escape the sandbox to leak your files. It just needs to run curl. We looked inside Grok Bot and Muse to answer one question: what happens after the injection lands? Our take: command review isn't enough. The check belongs at the exit, run by something the agent can't turn off.
12
10
3
87
426,711