zicato-evolve

Multi-agent evolution loops are automated by driving a self-improving harness through proposal, tournament evaluation, and promotion rounds across configurable generations. This enforces budget controls, epoch management, and safety gates to prevent ad hoc orchestration of agent experiments.

Agent tournament evaluation is a process where revised agent instructions are compared across a board of tasks to promote improvements that reduce loss. You need it when running self-improving multi-agent systems and evaluating optimization quality through structured logs.

Real-model execution in a self-improving agent harness is safely run by enforcing explicit approval before real-LLM runs. You can configure wall-clock limits, rejection thresholds, and use deterministic mock smoke tests to validate the evolution loop before live execution.

Yes, you can monitor tournament optimization quality by launching the live dashboard, which reports its URL and provides health checks and structured logs. This allows you to diagnose the multi-agent system's performance during the evolution loop.

Yes, you need separate harness and auxiliary LLM callables to run agent evolution rounds. Additionally, a configured zicato workspace and explicit approval for real-model execution are required to start the optimization loop.

Without budget controls like wall-clock limits and rejection thresholds, self-improving agent experiments face unbounded resource consumption and unmanaged epoch handling. Configuring these operational guardrails prevents runaway optimization and ensures safe tournament evaluation.