Whether an agent can actually do the work starts with the harness layer

Rohit Ghumare has recently been pushing a term: Harness Engineering. In his breakdown, an agent that can actually get work done has to complete seven steps:

• Load the right code, documentation, and history
• Let the model propose the next action
• Check permissions before running tools
• Execute within time and cost limits
• Verify results with tests and policy checks
• Feed failures back into a bounded retry loop
• Keep logs and traces for debugging

These seven steps sound like engineering common sense. In an agent setting, missing even one step causes problems.

The harness is the execution skeleton outside the model. It determines what context the model sees, which tools it can call, and when it must stop. Rohit's definition is straightforward: the harness defines what an agent can do, how work gets checked, and when it must stop.

The flowchart in his tweet illustrates this chain more clearly: the user goal 'fix the failing tests' first enters Context, which contains code, documentation, and history; the model proposes tool calls; the permission gate decides whether to allow or block them; the tool execution layer includes examples such as GitHub and Docker; finally it reaches Verify, which checks tests, policies, and limits. Three control items hang on the right: Permissions, Time + budget, and Logs + traces.

Many agent demos only capture the two stages of 'model proposal' and 'tool execution.' Permissions, budget, verification, and logging are omitted. The demo can run, but production doesn't dare use it. The reason is not complicated: without a permission gate, the model calling the wrong tool is an incident; without a budget, a loop can burn until the bill explodes; without verification, whether the result is right is all guesswork; without logs, even reproducing a failure is impossible.

Rohit linked the AI Engineering from Scratch course in his tweet. The page says it has 542 lessons, 20 stages, four languages: Python, TypeScript, Rust, and Julia; algorithms are implemented from raw math first, then frameworks are introduced. It is free and open source. If you want to brush up on agent engineering and this harness layer, you can follow the link.

Course link: AI Engineering from Scratch

Original tweet: Rohit Ghumare

Whether an agent can fix bugs, run tests, and commit code on its own looks like a question of model capability on the surface, but in the system it depends on whether the harness is complete. If a few of the seven steps are missing, you'll eventually have to learn the missing lessons in production.