Optimise cost per unit of work, not cost per token.
A task costs $0.02 on the small model and $0.20 on the large one — but the small model needed three attempts and the third was subtly wrong. Token price is a measurable proxy for the wrong thing.
The Agent Execution Optimization Engine routes on what actually happened in your workload: completions rather than tokens, efficiency tracked per task type, and budget enforcement that stops runaway retries.
Not in the request path · every routing decision a commit
Source-available under BSL-1.1, converting to Apache 2.0 18 months after each version's first publication. Internal use — including in production — is unrestricted.
What git was to source code, AgentStateGraph is to agent state.
Write a fact once. Every tool remembers it.
The context lives with the code, not in a chat log.
A conversation is a tree. Stop storing it as a line.
One passing run is a sample, not a test.