Auto Agent Long-Task Engineering Framework

Problem Being Addressed

Long coding tasks often span multiple files and are prone to context pollution: a model that implements and reviews its own work can mistake assumptions for facts and miss its own errors. Auto Agent reduces this by decomposing work into small, verifiable tasks and separating roles.

How It Works

• Role separation: Planner breaks down tasks and analyzes architecture; Generator implements and self-tests; Evaluator independently reviews code against the contract.
• Sprint Contract: each task writes sprint-contract.md with goals, acceptance criteria, target files, definition of done, and test method.
• Task decomposition: each atomic task touches 1-3 files, and tasks.json tracks dependencies, acceptance_criteria, steps, and complexity.
• Evaluation matrix: scores functional correctness, architecture compliance, code quality, and reusability; any hard threshold failure sends the work back for fixes.
• Session recovery: progress.md, handoff.md, and code files restore context; three failed iterations mark the task blocked.

Boundaries and Notes

Use it for large refactors, new modules, new systems, or explicit long-running work; avoid single-file fixes, queries, and simple bugs. It expects separate git commit for each task and human acceptance by default at Level 1. Failures must not be reported as passing.