patforna/auto-task
9 stars · Last commit 2026-09-08
An opinionated agentic workflow to turn well-defined units of work into shippable code with minimal human input.
README preview
# auto-task An opinionated agentic workflow to turn well-defined units of work into shippable code with minimal human input. **Think cruise control, not Full Self-Driving.** _auto-task_ aims to deliver high-quality code with high autonomy. To achieve this, it combines human-curated skills with research-backed techniques: workspace isolation, persistent task state, fresh-context subagents, multi-model-family panels and synthesis, TDD, evidence-cited triage, and more. Built for experienced engineers who want to ship software at an accelerated pace without giving up control. <p align="center"> <img src="demo/demo.gif" width="738" alt="A terminal: /at:create-task turns a one-line request into task 001, asking one question along the way. /at:auto-task then runs preflight, planning, implementation, review and verification unattended, and reports back for signoff: 8 commits, 3/3 tests passing, 1 flag, 9m 20s. A human answers "yes" and it ships."> <br> <sub><em>Marketing version</em> of an auto-task run. For a real run, see <a href="examples/runs/083-persist-column-picker.md">this sample transcript</a>.</sub> </p> > [!NOTE] > This is a very early extract from my daily workflow — rough edges included. Feedback, Issues and PRs welcome! For the story behind it, see [Getting Out of the Loop: 8 Months Solo-Building with AI](https://github.com/patforna/writing). ## How It Works