Blog
Engineering, product, and security.
Thoughts on AI agent infrastructure, self-hosted tooling, and building ClusterCode.
Prototyping — race your favorite agents on the same brief
Model choice is usually a debate: someone swears by Claude, someone else by Codex, someone points at a Copilot benchmark. Prototyping — in Beta today, on every plan — settles it by running the experiment. One brief fans out to two or more engine-and-model contenders, each works the task autonomously on its own branch, and you compare the finished results side by side. Promote the one you like and it opens the pull request.
Javier Perez · 2026-07-13
Introducing Loops: Run Until Verified, Checked by a Model That Did Not Write the Code
A month ago we wrote that ClusterCode was most of the way to a loop-engineering platform, and that the missing piece was a goal that keeps iterating until a separate model confirms it is genuinely done. We built that piece. Here is how Loops works — the maker/checker split, the gates, the guardrails, and why the last step is always a pull request you review, never a merge.
Javier Perez · 2026-07-06
Case Study: A Feature Delivered Across Jira, Confluence, and Bitbucket — With None of Their CLIs Installed
We gave an autonomous agent a Jira issue key and a high-level prompt — no tool, provider, or language named — and it installed the CLIs it needed, read a Confluence spec, ran a Claude-and-Codex review loop over a FastAPI repo, and opened a Bitbucket pull request via REST. Here is the trace, including the one rough edge.
Javier Perez · 2026-06-27
Loop Engineering: Stop Prompting, Start Designing the Loop
The field converged on a new idea in a single week: stop prompting your coding agent and start designing the system that prompts it. Here is what loop engineering means — and why ClusterCode is already most of the way there.
Javier Perez · 2026-06-24
Introducing ClusterCode: Self-Hosted AI Coding Orchestration
Today we are launching ClusterCode — a platform that turns your own machines into an AI development fleet. Paste a ticket URL, and a DevBox spins up with your AI agent cloned in, context loaded, and ready to code.
Javier Perez · 2026-03-15
Why We Built ClusterCode: The Problem with Running AI Agents Unattended
Every major AI coding tool ships with a scary-sounding flag to disable safety prompts. We asked: what if the environment itself was safe, so the flags did not matter?
Javier Perez · 2026-03-17
Getting Started: Your First AI Agent in 5 Minutes
A step-by-step guide to setting up ClusterCode, registering your first worker, and launching an AI coding agent on a real GitHub issue.
Javier Perez · 2026-03-20
Windows DevBoxes That Boot in Seconds, Not Half an Hour
ClusterCode runs Windows as a full VM, so a cold boot means a 20–30 minute Windows install before you can even start installing tools. A two-level golden-image recipe turns that into a seconds-long launch with your entire toolchain baked in.
Javier Perez · 2026-07-01
Ready to run AI agents in isolated DevBoxes on your own machines?