Agentelligence

Guidance for longer coding jobs

The entry skill tells Codex which guidance to use. Codex does the work with the project’s tools and instructions; extra workers and handoffs are optional.

The plugin has a local developer preview. I have not released it publicly.

Checking what changed

Agentelligence asks Codex to read current project sources and reuse settled facts until relevant inputs change. For delegated work, each worker gets a defined scope and returns what changed, what it checked, and what remains unresolved.

When a worker reports completion, the check goes back to the work itself. That might mean inspecting a code change, reading a test result, opening a generated file, or checking a deployment. Each needs different evidence.

Choosing where to spend effort

Extra workers and review cost time and tokens. I use the task's complexity and risk to choose the model, reasoning level, skills, and checks.

When I compare approaches, I record quality and usage together. Extra coordination has to justify its cost.

Recording the result

For prepared evaluations, I record the outcome, evidence, limits, and any quality or usage measures being compared. Those records describe the tested tasks; they do not prove that every task benefits from the plugin.

I also test fresh workers on prepared tasks with the pass conditions hidden. That gives me a separate basis for judging the result.

The screenshot shows an earlier roadmap dashboard. Its counts belong to that snapshot.