Summary
David Ondrej shares his agentic engineering workflow after more than 15,700 sessions and over 2,500 hours of coding with AI. He covers which models and harnesses he relies on, which subscriptions give the best value, how he reviews code to ship to production with confidence, and how he uses loops to automate parts of his daily work.
His current picks put Opus 5.5 at the top, with OpenAI’s GPT6 Astra as the strongest alternative, and he discusses how Google, Meta, and Kimi models fit in below the frontier. He argues that harness choice matters less and less as models gain terminal and internet access, citing conversations with Anthropic engineers. On subscriptions, he rates the Claude $100 and $200 plans as the best value, calls OpenAI’s recent plan changes inconsistent, and recommends OpenCode Go as a $10 option for tighter budgets.
A large section focuses on cloud agents versus local agents. He argues that most work should run in the cloud, and that people who prefer local agents are usually reacting to poorly configured cloud setups with missing skills, API keys, or dependencies. He demonstrates his open-source tool Cloudroom, including configurable permissions for how much access cloud agents get to a local machine, and explains when to use sub-agents versus cloud agents.
📺 Source: David Ondrej · Published October 30, 2026
🏷️ Format: Workflow Case Study
