Something interesting happened in AI this month.
On August 11, OpenAI deprecated custom prompts in Codex and told users to move reusable instructions into Skills. Claude Code now comes with skills such as /debug, /code-review, /batch, /loop and /verify. Google’s ADK 2.0 has moved to a graph-based workflow engine. And the July 28 MCP release added a stateless architecture, long-running Tasks and even Skills over MCP.
Then, on August 13, Google released Gemini 3.7 Flash specifically as a new workhorse for coding and agents. OpenAI is now pushing GPT-5.6 across its agent stack, while Anthropic has Fable 5, Opus 5 and Sonnet 5 built heavily around agentic work.
These look like separate updates. They are really pointing in the same direction.
A prompt tells AI what you want. A harness gives AI a way to work.
That difference is becoming one of the most useful AI skills to understand.
A good model can think. A good harness can make that thinking repeatable.
So, what exactly is a harness?
Forget the technical definition for a minute.
Imagine you hire a very smart person.
If you put them in an empty room and give them one instruction, their result depends almost completely on what they already know.
