Multi-Engine Coding Agent Setup: Routing Tasks Across Claude, Codex, and Ollama
Ok_Shoulder9804 · reddit · 2026-08-08
The author shares a multi-engine agentic coding harness built for a 15-person team. The core design philosophy is refusing to lock into a single provider, routing tasks through a unified session abstraction to Claude Code, Codex, and Ollama.
The system supports dynamic model selection based on task value: cheap models handle simple variable renames or config tweaks, frontier models tackle multi-file refactors, and sensitive data stays local. Adding a new model is just a config entry.
The main pain point highlighted is that the three CLIs agree on almost nothing—session resumption, streaming formats, and approval prompts are entirely different. Roughly 80% of the effort went into normalizing these into one interface. The setup has processed 1,851 sessions so far.
More from coding & agent
- Google DeepMind Expands CoT Monitoring to All Astra Agentic Applications — dfrsrchtwts · 2026-08-08
- Magnitude: An Open Source Local AI Agent with Built-in Inference Engine — ycombinator · 2026-08-08
- Disabling Auto-Review Agent Significantly Saves AI Coding Credits — carlosdponx · 2026-08-08
- Microsoft Launches MCP C# SDK v2.0 with Stateless Architecture — unixterminal · 2026-08-08
- NVIDIA Releases NeMo Gym Conversational Tool-Use Assets — _akhaliq · 2026-08-08
- Zero-Model Screen Compilation: 86x Agent Memory Compression at 98% Accuracy — Nossa Iyamu · 2026-08-08