To unlock the holy grail of AI, setting up feedback loop becomes critical. This is irrespective of how you are building your AI workflow,
- using an existing harness like Claude Code/Co-Work, Cursor, CoPilot
- building you own agent
Set up an coded/scripted feedback, that can be invoked by the ai agents, for them to validate if their work is correct.
In coding context, ensure entire code lint, architecture validation, unit test, integration tests, UI test like web/mobile, api tests, contract, can run where the agent is running.
This can be achieved with harnesses like Claude Code, using hooks, loops, skills along the scripts that run these exit criteria.
We will have to make the verification output more accessible by AI. For example, test coverage reports are generally configured as html files for human conception. But a json or XML, along with a quick parsing scripts added to the workflow will allow the agent to make sense, where the coverage is missing, in a repository of million lines spread across hundreds of file.
Comments
Post a Comment