Skip to content

Instantly share code, notes, and snippets.

@itissid
Last active February 8, 2026 00:03
Show Gist options
  • Select an option

  • Save itissid/cb0a68b3df72f2d46746f3ba2ee74afe to your computer and use it in GitHub Desktop.

Select an option

Save itissid/cb0a68b3df72f2d46746f3ba2ee74afe to your computer and use it in GitHub Desktop.
```
(/research -> agent -> research_doc -> (parsing to struct -> agent_run -> serialized_json) -> human_in_loop_verify)
|
| ^
research.md
# Task: Make a sandwich |
|
## User Story |
... |
## Instructions |
... |
## Progress Log |
... |
## Review Decision |
... |
struct: {
"gap": [Gap]
"technical_approach_solving": [Approach]
}
Gap analyzer Agent
System Prompt: "you are an expert at analyzing gaps in the code base,
extract all the gaps from research doc and add
implementation ideas in above format: .... Output
in structured JSON. Present to human for verification
focusing on "
```
@itissid
Copy link
Author

itissid commented Feb 8, 2026

In this toy example — just research —  here the idea you start with a very customized doc, e.g. research.md. Then the Gap analyzer Agent which is essentially like your very familiar, sub agent, creates structured data from the parsed document. Why? Because you need to not give the next step — a human — the entire document, just the important stuff. The presumption here ofc is that presenting gaps is the timely and appropriate thing to do, its a design decision by the person making these complex beasts.

These steps can be more complex and repeated multiple times a day, e.g. an AI could arbitrate things and ask more humans about the gaps, because everyone is too busy to look at it, multiple rounds of feedback from humans etc.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment