synth-ai 0.9.3

Serverless Posttraining for Agents - Core AI functionality and tracing

Synth

Build systems for OOMs more complexity.

Continual and offline optimization for prompts, context, skills, and long-horizon memory.

Use the SDK in Python (uv add synth-ai) and Rust (beta) (cargo add synth-ai), or call Synth endpoints from any language.

Synth Style

Synth is built for frontier builders first. We:

• push interface complexity inward (strong server contracts, simpler app surfaces)
• design online/offline parity with pause/resume as first-class controls
• meet production code where it is (no forced lock-in or rewrites)
• build general algorithmic foundations, then layer targeted affordances

For engineering principles and coding standards, see specs/README.md.

Average accuracy on LangProBe prompt optimization benchmarks.

Demo Walkthroughs

• GEPA Banking77 Prompt Optimization
• GEPA Crafter VLM Verifier Optimization
• GraphGen Image Style Matching

Benchmark and demo runner source files live in the Benchmarking repo (../Benchmarking in a sibling checkout).

Product Focus

• Continual Learning Sessions (MIPRO + GEPA): run online sessions that update prompts from reward feedback during live traffic, with first-class pause/resume/cancel controls.
• Discrete GEPA Optimization (Prompt + Context): run offline GEPA jobs for controlled batch optimization, compare artifacts, and promote the best candidates.
• Voyager for Skills + Long-Term Memory: optimize skill/context surfaces and use durable memory with retrieval and summarization for long-horizon agent systems.
• One Canonical Runtime Surface: use shared systems, offline, and online primitives across SDK and HTTP APIs.
• Agent Infrastructure Built In: run with pools, containers, and tunnels for local or managed rollouts without forcing app rewrites.
• Graph + Verifier Workflows: train GraphGen pipelines and rubric-based verifiers for domain-specific evaluation loops.

Getting Started

Python SDK

uv add synth-ai
# or
pip install synth-ai==0.9.3

Rust SDK (beta)

cargo add synth-ai

API (any language)

Use your SYNTH_API_KEY and call Synth HTTP endpoints directly.

Docs: docs.usesynth.ai