Fork Everything
We replaced a 200K-star framework with 5,000 lines of Python. Agentic coding killed the one argument against forking.
Don’t fork open source. Contribute upstream. You can’t maintain a fork.
That was good advice. Frameworks gave you hundreds of contributors for free. Every bug fix, security patch, and new feature showed up without you lifting a finger. Sure, you carried a ton of code you never used, but who cared? You weren’t the one maintaining it. Free labor, free improvements, free ride.
Agentic coding changed the math. When Claude Code can build exactly what you need in an afternoon, all that extra code stops being free. It’s noise your AI has to read, abstractions it has to navigate, complexity that slows down every change you make. Lean codebases are a competitive advantage now. And building your own just got trivially cheap.
200K Stars, 5% Useful
OpenClaw has 200K+ stars. Its creator just joined OpenAI. It handles agent orchestration, skill registries, tool management, context handling, evaluation. We were using 5% of it.
That other 95% wasn’t free. It was attack surface, abstraction layers, and hours spent debugging someone else’s design decisions. Every big open source project works this way. You adopt the whole thing to use a fraction of it.
What We Actually Needed
We stripped OpenClaw down to the agent runtime loop and called it Pinecone. Pure Python, no Node.js dependency, no framework abstractions. ~5,000 lines that do exactly what our agents need.
Then we built what OpenClaw doesn’t have: an affinity matrix for agent-task matching, state validation, a custom eval framework, context truncation tuned to our workloads, a permission model designed for our threat model. 9 agents run 24/7 on it.
We wanted full visibility into the agent fleet too. Real-time dashboard showing what every agent is doing, token burn rates, cron health. We could have used OpenClaw’s built-in tooling or bolted monitoring onto the existing stack. Built our own instead. FastAPI backend, Alpine.js frontend, SSE updates every 5 seconds, kanban board for agent tasks, fleet overview, token tracking, cron monitoring. Days of work, not months. Claude built most of it.
Lean Codebases Win Now
With OpenClaw, Claude Code scans your project and most of what it reads is framework code you never touch. Irrelevant abstractions, features for use cases that aren’t yours, architecture decisions someone else made for someone else’s problems. Claude has to parse all of it to find the parts that matter.
With Pinecone, every line is ours. Claude understands all of it. Ask it to change how agents handle context truncation and it doesn’t navigate three layers of abstraction to find the right hook. It just changes the code. Your codebase is 100% relevant to what you’re actually building, and Claude is faster and more accurate because of it.
Research on Demand
When we need something new, we don’t read OpenClaw’s source. We tell Claude: “research how agent frameworks handle X, then build a version for Pinecone.”
We built an entire LLM eval framework this way. 12 files, recall testing at different context utilization levels, Claude as impartial judge. Claude researched eval patterns across the ecosystem and implemented a custom version tuned to our agents. Same approach for agent feedback loops. “Find solutions others have implemented for these failure modes” produced a custom scoring system with a penalty matrix.
Maintenance was always what killed forks. Keeping up with upstream, porting security patches, adapting to breaking changes. Claude handles that now. And because your codebase is lean, Claude handles it well. No wading through framework internals to understand where a patch applies. No deciphering abstractions you didn’t write. Just your code, clearly scoped, easy to modify.
Personalization All the Way Down
Same principle from Where the Work Goes: when building gets cheap enough, generic stops making sense. You don’t hyper-personalize your product for users and then run your stack on someone else’s generic framework. Customization goes all the way down.
If you’re running a framework and using 5% of it, you’re carrying dead weight because someone told you that you couldn’t afford to build your own. You can now.