-
Notifications
You must be signed in to change notification settings - Fork 0
Glossary
A system that helps decide what an agent should do next, including planning, orchestration, tool loops, memory, user interaction, and application control flow.
A provider-specific or gateway-specific implementation that lowers an optimized mIR plan into concrete execution.
A versioned description of provider or backend features, such as prompt caching, structured outputs, batching, tool calling, retention options, and context limits.
A first-class mIR artifact representing a unit of context such as a system instruction, user input, retrieved document, memory, tool result, example, scratchpad, or validator output.
The artifact Migaki emits to show what changed, why it was allowed, what assumptions were used, and how the result can be audited or replayed.
A graph of model calls, tool calls, retrieval calls, context artifacts, validators, branches, joins, retries, fallbacks, approvals, and constraints.
A system that transforms a logical execution graph into a measurable execution plan under declared constraints.
The process of translating a portable mIR plan into provider-specific, gateway-specific, or runtime-specific execution instructions.
Migaki Intermediate Representation, the portable logical representation of agentic execution.
A transformation over mIR that produces a new plan, a plan diff, evidence, and warnings.
Describes logical plans that are not tied to a single provider API shape.
An anti-goal. A provider-naive system ignores real provider differences and loses important execution capabilities.
The idea that an optimized execution is acceptable when it preserves the declared task outcome under validators, evals, invariants, and allowed regression thresholds. It does not mean byte-identical or semantically identical model output.