Design Your Own Coding Agent Harness

The first lesson from the open-source course Building a Coding Agent From Scratch: https://github.com/decodingai-magazin... Complementary articles: Coding Harness Architecture: https://www.decodingai.com/p/building... The Bare-Bones Coding Agent Loop: https://www.decodingai.com/p/the-codi... --- A design walkthrough of a coding agent harness built from scratch: the AI agent loop, the 5 harness modules around it, and the 2 interface modes on top (terminal and remote). You get the full architecture here, ready to adapt to your own AI application. You'll walk away knowing how to: Build a coding agent loop on Pydantic AI with 4 core tools: read, write, edit, and bash Choose an LLM provider on the build vs buy axis, from proprietary APIs to open weights you serve yourself via Modal Do the cost math before you build, comparing pay-per-token pricing against GPU time, and serverless against reserved GPUs Draw a clean line between the agent loop and the harness modules around it: sandboxing, memory, compaction, skills, and an LSP server Run one harness in 2 modes: an interactive TUI fed by a steering queue, and a remote headless mode on Kitaru with durability, checkpoints, replays, and scaling Place AI evals via Opik on top of the harness and know which question each one answers: internal benchmarks, regression tests, and production evals For: AI engineers who use tools like Claude Code or Codex every day and want to know exactly what happens under the hood, whether to become power users or to build their own harness. Tech stack: Modal Endpoints (open-weights inference): https://tinyurl.com/2t83yyxp Modal Sandboxes: https://tinyurl.com/89sr426z Opik (observability & AI evals): https://tinyurl.com/mssd4ufp Kitaru (durability & AI evals replays): https://tinyurl.com/5n8vvzcc Pydantic AI (agent loop): https://tinyurl.com/yuyxjrjn Speaker info: LinkedIn:   / pauliusztin   X: https://x.com/pauliusztin_ Blog: https://www.decodingai.com/ Timestamps: 00:00:00 - Intro 00:00:40 - Demo: the coding agent we're building 00:02:52 - The 6 demo skills 00:04:52 - System design overview 00:07:04 - The agent loop and the 4 core tools 00:09:29 - The 5 harness modules 00:12:36 - LLM providers, build vs buy 00:14:03 - Cost: per-token vs GPU time 00:17:23 - Serverless vs reserved GPUs 00:20:01 - Interactive mode and the steering queue 00:21:50 - Remote mode, durability, and replay 00:24:31 - Deployment architecture 00:25:56 - Why only 2 interface modes 00:26:57 - Observability: threads, traces, spans 00:28:39 - AI evals: benchmarks, regressions, production 00:32:13 - The course repo 00:33:54 - Running the knowledge graph demo 00:37:01 - Debugging the run in Opik