Agentic Coding Harnesses
This post extends the concept of agentic coding harnesses by detailing strategies to overcome rate limiting issues with powerful LLMs and leverage past conversation data for continuous improvement. It demonstrates how to build and deploy applications by intelligently mixing different models and providers, using smaller, cheaper models for less critical tasks and reserving high-end models for planning and review. The author showcases an 'AI Software Factory' and provides a 'build-dark-factory' sk. This post further elaborates on agentic applications by showcasing how large companies are deploying AI agents not as single-agent interactions, but as multi-agent systems orchestrated within workflows. It highlights the use of AI only where judgment is required, the translation of rules into tested code, and the necessity of human approval for critical decisions. The Oracle AI for Fusion Applications is presented as a case study, demonstrating agents that prioritize actions and execute every LLM call within a workflow, incorporating security policies and visual outputs instead of raw text.
Timeline
