Mahal Systems The Agent Optimization Toolkit
Make your agents 1% better every day with AOT-enabled evals and optimization strategies.
local maxθ* How it works
Build a flywheel around your agent development process with AOT.
Q(θ)Q*(θ) Measurement
Cost, quality and latency need a baseline that holds before anything can move.
What we built: End-to-end creation of high-quality evals
AOT builds your agent registry and living evals from real traces: cases that find errors, a set sampled by traffic share to estimate quality, AI judges calibrated to your team's labels, and repeated runs that set the noise floor. As usage shifts and the agent evolves, the evals are re-cut to match.
todayθ* Iteration method
Vetted strategies, each with its implementation and testing strategy.
What we built: A complete & vetted catalog of optimization best practices
AOT catalogs the latest agent optimization methods, 60 and counting, from model routing to tool design. Each is cataloged by the workload it suits, by what it changes (models, tools, prompts, context, memory, orchestration, retrieval, verification), by its expected effect on quality, cost and speed, and by the strength of the evidence behind it.
Qt →with AOT Compounding asset
Every test, trace and eval result aggregated into a central database that measures what works and what doesn't.
What we built: Our optimization inference engine
The software
A customizable open-source package that deploys into your production environment. The plugin manages the process and interfaces with your coding environment; the engine stores the index and drives optimization strategies.
Your infrastructure
Principles
Deployment
AOT is free. If you need help designing the architecture, the testing environment, the performance index, or anything in between, we'd love to help.
Need help?