Rafa Jiménez

💯 that's the reason why we are working on Kyoei, bring the human back at the center of AI-assisted research and engineering. Repo + paper: github.com/sinequanone-org/k… Video: youtu.be/UMe3w_u-Ka4

#CaptainCode ❤️ #OpenShell. Ensure your prompt runs safely! CaptainCode has experimental support for @NVIDIAAI / @Canonical OpenShell. You define the task, permitted files and tests. The worker runs inside a sandbox with policy-controlled file and network access. Tests run in the sandbox, and Captain Code returns a checked patch for review. 💡 Quick example of how this helps. Imagine you want to task your agent with improving a cleanup script on your laptop, and you are worried it might delete the wrong files during its tests. You absolutely want to test that safely in a sandbox! 👉 Just type in CaptainCode: /openshell test and improve scripts/cleanup.py, making sure it only deletes *.md files ⚙️ Captain code will confirm: allowed=scripts/cleanup.py verify=["python3","-m","pytest","tests/test_cleanup.py"] and come back with a verified patch and test results, all from a disposable sandbox. 🔐 Nothing ran on your machine. You review the diff, confirm the tests pass, and apply it only when you are ready. ✅ What lands in your terminal: • Snapshot taken at HEAD • Sandbox worker 1/1: scripts/cleanup.py + tests/test_cleanup.py • Verify: pytest tests/test_cleanup.py — 3 passed, 0 failed • Advisory host review: no issues found • Patch exported: /tmp/cc-openshell-5a9f.patch Thank you @jtm_tech, @AlexWatson405, @a_golshan @NVIDIAAIDrew and the broader community for bringing OpenShell to life 👏 We intend to contribute boundary tests, runtime fixes and evidence formats to OpenShell. 📖 Learn more: lemma.ventures/blog/captain-… 🕹️ Test it today: captaincode.ai 🧐 Release notes: github.com/lemma-ventures/ca… 🐚 OpenShell: github.com/lemma-ventures/ca… #AIAgents #DevTools #OpenSource

Same AI/LLM prompt, with different answer, leads to a different set of decisions. That is the failure mode regulators and auditors actually care about when AI agents move money, unlock assets, or change who can act. Most teams still hear "the model is non-deterministic" and stop there. In practice, the model weights are often stable. What drifts is the serving stack around them: hardware, engine build, scheduler, silent updates. Temperature 0 does not mean replayable. For regulated and mission-critical systems, three questions come first: 1. Can we reproduce the same output from the same input? 2. Can we name what changed when the output changed? 3. Can we prove the path met policy at the time it ran? If the answer is no, you do not have an audit trail. You have a plausible story, not a verifiable, provable, replayable one. I wrote a short, non-technical note on why replayability is a foundation for agent adoption in regulated systems, and how to ask for it before launch. Read it here: lemma.ventures/blog/regulate… Disclosure: I maintain the harness under Lemma Ventures AG, which builds infrastructure in this space. Read incentives accordingly. #AIAudit #RegTech #AIAssurance #AgenticAI