I built AEGIS for software requests that did not contain enough information to begin work safely. A request could omit its scope, required dependencies, or failure conditions. The planning stage identifies these gaps before execution begins.
Autonomous software delivery depends on separating work that requires judgment from work that requires execution. AEGIS uses stronger models to plan unclear work. It sends defined tasks to lower-cost models or software processes with the required capability.
AEGIS accepts a goal through Telegram or a local operations interface. One message can start the design, build, test, and deployment of a web app.
The system defines research, writing, software, image, video, and audio tasks before assigning them. Each task has a purpose, required inputs, required output, and acceptance test.
AEGIS runs independent tasks at the same time, up to a set limit. A dependent task begins only after its required input is complete. The task definitions keep concurrent work controlled and repeatable.
AEGIS stores every state change and task result in the project folder. It also stores command results, file changes, reports, and audit records.
An independent review process checks the complete result against the original goal and stored evidence. AEGIS stops when the review process cannot make a valid decision. Completing every task does not mean the complete request is ready for delivery.
AEGIS saves checkpoints at important stages. If a task fails or an audit finds a problem, the system isolates and changes only the affected work. It then runs the audit again.
I added cost limits, model alternatives, run search, schedules, and approval controls. AEGIS requires approval before using credentials, adding a paid tool, or increasing a budget.