Engineering Notes
Notes from building OpenOrc, newest first. Also available as anRSS feed.
- A Group Chat Is Not an Engineering Workflow
Slack Code makes AI-assisted coding collaborative. OpenOrc takes a contract-driven approach to autonomous engineering.
- Stop Sharing Context Between AI Agents. Start Sharing Contracts.
How we're building an AI-native engineering workflow around specialized roles, durable knowledge, independent review, and human accountability.
- AGENTS.md at Scale: A Practical Guide to Nested Instructions
How to structure nested AGENTS.md files so coding agents respect architectural boundaries, run the right tests, and avoid unnecessary context.
- The Danger of Losing Focus
How a chain of reasonable infrastructure decisions nearly pulled OpenOrc away from its actual objective, and why architectural boundaries matter before the compiler can see them.
- Stop Managing AI Agents Like Employees
AI agents can perform remarkable amounts of engineering work without becoming employees. OpenOrc keeps agent provenance inside the workflow while leaving accountability with the human engineer.
- Why AI Review Needs a Stopping Rule
Independent AI review catches real engineering mistakes. It can also keep inventing increasingly remote reasons not to approve the work. A useful Reviewer needs an explicit definition of good enough.
- "Implement Account Deletion" Cost 100 Million Tokens
A seemingly ordinary OpenOrc backend task consumed more than 100 million Cline tokens. The interesting part was not the number, but the engineering and inference economics hidden behind the issue title.
- OpenOrc Started as Copy and Paste Between Two AI Models
The producer-reviewer workflow behind OpenOrc began as a manual loop between two independent AI systems that consistently produced better software together than either did alone.