Multi-agent systems typically fix their team structure up front. This report presents an architecture in which the system governs itself: an expert role roster is resized dynamically against task complexity, execution profiles are selected by a threshold-based Thermal Escape and Threshold Decay (TETD) consensus vote among the agents, and work is executed through role-augmented context prompts.
The architecture is fully implemented as an open-source, multi-tenant Python
package (absolute-self-governance, MIT-licensed, on PyPI), and the report
evaluates it across three reproducible tiers — a 573-case deterministic test
suite, a live-infrastructure deployment, and live LLM API runs. The results are
reported without embellishment: the allocation model scales to 50 million
simulated roles in 0.084 seconds, but on routine coding tasks the full
role-informed pipeline showed no aggregate quality advantage over a
single-agent baseline while costing 3.6× more — evidence that role-prompt
overhead should be reserved for complex control logic rather than routine
tasks.