We’ve been using Claude Code every single day for the past year. Organizing files on our computer. Building custom web apps. Modeling in Blender through the browser. Color grading footage. You name it, we’ve pointed Claude at it. We try to maximize every tool Anthropic ships, because the leverage compounds when you actually live inside these tools instead of dabbling.
Check it out for yourself here.
But coding was different. We kept hitting the same walls. Plan mode would get us 80% of the way there and then run out of room. Each session was isolated, no memory of what came before. Context ceilings would force us to chunk work into pieces small enough to fit, which meant losing the threading between decisions. We’d spend more time managing Claude than building.
So we went deep. We spent months researching progressive disclosure, context architecture, agent orchestration patterns, and how the Anthropic stack actually works under the hood. We rebuilt our internal workflow from the ground up around a single idea: if you front-load the discipline of a real product requirements process, the agents can run for hours without losing the plot.
That became KARIMO. It’s our internal harness for Claude Code, and we use it every day at Open Session.
What it actually does
KARIMO turns a feature idea into a structured PRD through a guided interview, decomposes that PRD into isolated task briefs with full context, and then dispatches a fleet of agents to execute those tasks in parallel git worktrees. Each agent has its own branch, its own scope, and its own review loop. You walk away from your computer, and a few hours later you come back to a feature branch with a real diff, a real PR, and a real audit trail.
Here’s what makes it different.
PRD-Driven. Most agent frameworks start with a prompt. KARIMO starts with a 10-minute structured interview that captures requirements, scans your codebase for relevant patterns, and produces a document that’s optimized for agent comprehension. The agents are not guessing what you want. They have a spec.
Extended Planning. This is plan mode, but at a scale plan mode was never built for. Hundreds of tasks, executed with accuracy, with full lineage back to the PRD that requested them. You can trace every line of code to the requirement that produced it.
Context-Efficient. The framework is designed to use fewer tokens. Each task brief is self-contained, so agents only load what they need. No more burning context on irrelevant codebase chunks.
Configurable. You decide how it runs. Model routing (Sonnet for speed, Opus on escalation). Branching strategy. Commit behavior. Review loops. Fast track for trivial changes, full review for risky ones. It’s a harness, not a black box.
Fully Loaded. 22 agents. 11 commands. 8 skills. Coordination agents that orchestrate without writing code. Task agents that write production code, tests, and documentation. A reviewer that catches issues before they reach you.
This is not vaporware
We’ve been running KARIMO internally since the early experiments. The repo has over a thousand commits. We’re on version 9. We’ve used it to ship features across multiple web apps in production, and we’re using it right now to build BOS, our brand operating system platform.
To prove the harness works, we built a demo website using KARIMO end to end. PRD interview, task decomposition, parallel execution, review loops, the whole flow. The demo covers a real product surface with a working frontend, content architecture, and component library. It’s not a toy.
Why we’re sharing it
Honestly, because we think this is what Claude Code is supposed to feel like. The base experience is already incredible. The harness around it is what unlocks the next level. We don’t think we have the only good answer here, but we have a real one, and we’d rather put it in the world than sit on it.
KARIMO is open source under Apache 2.0. The repo has full documentation, a one-line installer, and a doctor command to diagnose any setup issues. If you have Claude Code, GitHub CLI, and a project to point it at, you can be running your first PRD in under five minutes.
We believe this harness only gets better as the models evolve. And we believe coding agents work best when they’re given the same scaffolding a real product team would demand.
Check it out for yourself here.
Go build stuff.
