Da Vinci Recursive Improvement CAD Harness

This sensor-mount study

1. Set the design contract. A Python runner fixes the mounting footprint, PA12 material assumptions, 20 N load and acceptance limits. Astra writes CadQuery wrappers and chooses dimensions within the supplied cradle family.
2. Build and measure. Candidate code runs in an isolated Docker container and exports STEP and GLB. A separate evaluator checks the STEP against the supported geometry family, measures volume and applies the same stress and deflection screen to every attempt.
3. Reflect and reuse. Previous results, scoped Atlas Vector Search matches and accumulated lessons enter the next prompt. After two attempts, Astra created a wall-thickness utility; five validation cases passed before the saved tool was reused.
4. Archive and select. Git versions source, tools and context. Atlas stores records and GridFS stores artifacts. The runner retains the lightest passing design; this page displays all eight attempts from an exported snapshot without making new model calls.

Full multi-agent workbench

The full harness coordinates structural and aerodynamic specialists against a shared vehicle specification. Two Python workers process durable jobs. Atlas candidate-insert triggers enqueue evaluation; evaluation-insert triggers enqueue reflection, with a recovery loop for missed delivery.

A meta-agent proposes reusable tools and versioned changes to orchestration logic, context policy and an isolated React policy panel. Independent tool tests and release checks run before activation. The evaluator and acceptance limits remain fixed outside those editable files.

The sensor study shown here uses a sequential Python loop with the same sandbox, archive and memory services; it does not use the full workbench’s two-specialist trigger queue.

Measurement scope

Mass uses STEP volume and nominal PA12 density. Stress and deflection are wall-strip estimates, not FEA. The VTOL is illustrative; measurements cover the mount only.