Ax (AxACE optimizer)

Type: full-code · Vendor: ax-llm · Language: TypeScript · License: Apache-2.0 · Status: active · Status in practice: emerging · First released: 2023-12-01

Links: homepage docs repo

Ax is a DSPy port that compiles typed declarative signatures into LLM calls and optimizes the resulting prompts and demonstrations against an evaluation metric.

Description. Ax is a framework for building LLM applications around declarative signatures rather than hand-written prompts. A signature declares the typed inputs and outputs of a step, and Ax turns it into the prompt, the output parser, the validators, and the retry loop, returning typed data. Ax includes optimizers, among them AxACE, which mirrors the Agentic Context Engineering paper's Generator-Reflector-Curator loop and a Genetic-Pareto optimizer that tunes prompts and demonstrations against evals. It is released under Apache 2.0.

Agent loop shape. A developer declares a signature stating what goes in and what comes out. Ax compiles the signature into a prompt, schema, parser, validators, and retry loop, runs it against the model, and returns typed output. Given example input/output pairs and a metric, an optimizer such as GEPA or AxACE searches for better prompts and demonstrations; AxACE additionally maintains a structured bullet playbook updated through add/update/remove deltas at inference time.

Primary use cases

building LLM pipelines from typed signatures
automatic prompt and few-shot optimization
agentic context engineering with an evolving playbook
type-safe structured outputs from LLMs

flowchart TD fw["Ax (AxACE optimizer)"] fw --> p1["DSPy Signatures (core)"] fw --> p2["Structured Output (core)"] fw --> p3["Agentic Context Engineering Playbook (first-class)"] fw --> p4["Tool Use (supported)"] fw --> p5["Agent-as-Tool Embedding (supported)"] fw --> p6["Multi-Model Routing (supported)"] fw --> p7["Streaming Typed Events (supported)"]

Key concepts

Signature → dspy-signatures (docs) — A one-line declaration of a step's typed inputs and outputs that Ax compiles into the prompt, output parser, validators, and retry loop, so the developer states what goes in and out rather than writing a prompt.
GEPA optimizer → evaluator-optimizer (docs) — The Genetic-Pareto optimizer that searches over prompts, demonstrations, flows, and agents against an evaluation metric to improve a signature-driven program without manual prompt edits.
AxACE → agentic-context-engineering-playbook (docs) — An optimizer that mirrors the Agentic Context Engineering Generator-Reflector-Curator loop, representing context as structured bullets and applying incremental add/update/remove deltas into a serialized playbook updatable at inference time.
AxAgent → agent-as-tool-embedding (docs) — A three-stage pipeline that turns a signature into a long-running, tool-using actor whose memory, skills, MCP tools, and child agents are exposed as typed capabilities behind the signature contract.
ai() provider handle → multi-model-routing (docs) — The single constructor that binds a program to a model provider — OpenAI, Claude, Gemini, a gateway, or a local OpenAI-compatible endpoint — leaving signatures, tools, traces, and outputs unchanged when the provider is swapped.