SciML author shares multi-model coding workflow: role-based routing for writing and review
ChrisRackauckas · x · 2026-11-18
Chris Rackauckas, SciML lead developer, describes how he routes multiple models by role in his coding workflow:
- Coordination & review: Carbon (strong at code review) is reserved for coordination; primary code review goes to Codex / Devin (Sol/SWE2 Fusion), with bulk workhorse review handled by OpenCode Go models (GLM 5.3, DeepSeek 4.1 Flash) and Grok 4.7.
- Code writing: Claude is preferred for writing; the new Hydrogen looks like good value; free SWE2 works surprisingly well when implementing plans Carbon makes, and being free, several can run in parallel.
- Spark scores too low as a reviewer and just wastes review time, so it's dropped.
The core idea: split coordination, writing, and review into separate roles and assign models by cost-effectiveness per role, letting free models handle parallelizable mechanical coding.
Related event: SciML Author Shares Role-Based Multi-Model Coding Workflow(2 posts)→
More from coding & agent
- Scientific computing lead shares multi-model coding workflow: free models write, Claude coordinates — ChrisRackauckas · 2026-11-18
- Capability Is Not Authority: The Hidden Risks of Deploying AI Agents — Adjust-the-rhythm · 2026-11-18
- AI Agents That Need Constant Babysitting Aren't Agents — Few-Garlic2725 · 2026-11-18
- Rebuilding the Metabase MCP Server: Lessons in MCP Tool Design — Ramirond · 2026-11-18
- Dev wants a single client to manage all agent threads with Slack-style swipe UI — curious_vii · 2026-11-18
- Adaptive UI in a Claude CLI Enterprise Assistant: Prompting Isn't Enough — Shivam__kumar · 2026-11-18