I built AEGIS for software requests that did not contain enough information to start work safely. A request could omit the scope, required dependencies, or failure conditions. I designed the planning step to identify these gaps before execution.
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.
Select a model for each task
AEGIS uses frontier models for tasks that require planning or judgment. It assigns defined execution tasks to lower-cost models that have the required capability.
The routing method reduces cost because AEGIS does not use the most expensive model for every task. Each task still uses a model or software process that can produce the required output.
Define each task before execution
AEGIS assigns research, writing, software, image, video, and audio tasks to suitable models or software processes. It runs independent tasks at the same time, up to a set limit. It does not start a dependent task until the required input is complete.
Each task has a purpose, required inputs, required output, and acceptance test. The task definitions make concurrent execution controlled and repeatable.
Verify the complete result
AEGIS stores each 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. A completed task does not mean that the complete request is ready for delivery.
Restart only the affected work
AEGIS saves checkpoints at important stages. If a task fails or an audit finds a problem, the system isolates the affected work. It changes only that work and runs the audit again.
I added cost limits, model alternatives, run search, schedules, and approval controls. AEGIS asks for approval before it uses credentials, adds a paid tool, or increases a budget.