Platform for building, evaluating, and optimizing AI agents in production. Provides structured event routing, long-term memory, prompt optimization, automated evaluation, and human-in-the-loop action approval.
| Crate | Version | Description |
|---|---|---|
horizons_core |
0.0.0 | Core domain — events, context refresh, agent actions, onboarding, project DB |
horizons_rs |
0.0.0 | HTTP API server (Axum) with dev-mode in-memory backends |
horizons_integrations |
0.0.0 | Connectors (Jira, LinkedIn, Langfuse) and queue backends (SQS, RabbitMQ) |
voyager |
0.1.0 | Agent memory — store, retrieve, and rank episodic and semantic memories |
mipro_v2 |
0.1.0 | Prompt optimization — dataset splits, candidate generation, early stopping |
rlm |
0.1.0 | Evaluation — reward signals, weighted scoring, pass/fail verification |
All crates use Rust edition 2024.
| SDK | Version | Path |
|---|---|---|
Python (horizons) |
0.0.6 | horizons_py/ |
TypeScript (@horizons/sdk) |
0.0.6 | horizons_ts/ |
Starts the Horizons server on http://localhost:8000 with persistent local storage. No external services required.
Verify:
curl http://localhost:8000/health
Prerequisites: Rust 1.85+ (edition 2024 support).
cargo build --release -p horizons_rs --features all
cargo run --release -p horizons_rs --features all -- serve
Starts on http://localhost:8000 with dev backends (SQLite + local filesystem). No external services required.
cd horizons_py
pip install -e .
cd horizons_ts
npm install
npm run build
horizons_rs (HTTP API)
├── horizons_core
│ ├── events — publish/subscribe on dot-delimited topics, glob matching, retry + DLQ
│ ├── context_refresh — pull from external sources on cron or event triggers
│ ├── core_agents — action proposals, risk levels, review policies (auto/AI/human)
│ └── onboard — project DB (Turso/Postgres/S3), user roles, audit log
├── horizons_integrations
│ ├── connectors — Jira, LinkedIn
│ ├── queue_backends — SQS, RabbitMQ
│ └── langfuse — trace export
├── voyager — episodic memory with relevance/recency/importance ranking
├── mipro_v2 — MiPRO prompt optimization with holdout evaluation
└── rlm — reward signals (exact match, contains, LLM rubric), weighted scoring
FSL-1.1-Apache-2.0 (the Sentry license) — Copyright 2026 Synth Incorporated.
Free to use, modify, and redistribute for any purpose except building a competing product. Converts to Apache 2.0 after two years.