This path is for you if a coding agent is a black box and you want the loop: tools, packed context, and how a run is judged. No part is live on this page yet. Start with AI agents for everyone for the plain version of watch, approve, and check.
- 1 What Is an AI Harness? The Loop Around the Model An AI harness wraps a stateless model with tools, state, and a ReAct loop. Build a small one yourself before you assume you need a proprietary agent product. Scheduled · November 16, 2026
- 2 Chat Box vs Agent vs IDE Agent vs Terminal Agent Choose chat for brainstorming, IDE agents for visual local edits, terminal harnesses for multi-file verify loops, and background agents for well-tested batch chores. Scheduled · November 17, 2026
- 3 Tools, Permissions, and ‘Approve Before It Runs’ Define tool schemas, intercept calls for human approval, and isolate execution. Host subprocesses need allowlists; containers need clear mounts. Defense in depth beats hope. Scheduled · November 18, 2026
- 4 Context Packing: Repos, Files, and Memory Pack agent context on purpose: search, AST skeletons, tiered budgets, and pruned tool output. More tokens are not more understanding. Scheduled · November 19, 2026
- 5 Evaluation and ‘Did It Actually Work?’ Evaluate coding agents with real issues, unit tests, and Pass@k metrics. Build a small internal golden set. Do not trust demos or pure LLM judges. Scheduled · November 20, 2026
- 6 Open Harnesses vs Product Harnesses Compare open and product harnesses on control, cost, and compliance. Prefer portable tool protocols so you are not stuck rewriting every connector. Scheduled · November 21, 2026
- 7 Safety: Runaway Loops, Secrets, and Destructive Commands Safety for agents means turn budgets, secret redaction, destructive-command gates, and hash-based loop detection. Assume the model will retry forever unless you stop it. Scheduled · November 22, 2026
- 8 When a Plain Chat Window Is Enough Use plain chat when the work is thinking and writing. Use agents when files, tests, and shells must close the loop. Match surface to stakes. Scheduled · November 23, 2026