Watch the loop before adding tools
To build an agent harness, first follow a single request through the software around the model. What conversation does the model receive? What action does it request? Which function runs that action, and how does its result become input to the next call? This short path is the foundation for understanding a larger agent.
Use a small teaching project
Paweł Huryn’s Mini Harness is an MIT-licensed Python teaching project, described upstream as roughly 1,400 lines. The HARNESS reel follows its 15-part diagram with source credit. Start by reading the source and tracing one simple task rather than treating a benchmark chart as a promise about your own workload.
Keep the first task disposable
Install with the upstream instructions, configure the model provider privately and start from a temporary project without secrets. First ask for a read-only explanation of a file. Inspect the requested operation and the returned evidence. Model calls can incur provider charges even though the harness code is free.
ONE TACTIC A WEEK
Separate permission from isolation
A permission gate decides whether an action may proceed. It does not automatically restrict what the process can access once that action runs. Inspect the tool runner and its environment as well as the approval prompt. Use an isolated environment when the task needs stronger boundaries.
Test interruption and completion
After a small, approved edit, review the diff and run a relevant check. Interrupt a session, resume it and inspect what state was retained. Ask whether the final result is supported by a test or an observable change. A confident completion message on its own is not evidence that the requested work succeeded.
Add one capability at a time
Memory, compaction, subagents, skills and external tools all change how a run behaves. Add one capability, then repeat the same bounded task so you can see what changed. Keep the useful traces and the failure cases. The companion guide explains the 15 parts and provides the installation command and first-run checklist.