Horizons OSS Launch, Environment Pools, RLM v1 & MIPRO Offline

TL;DR

Horizons launched as an open-source agent execution platform — context refresh, event sync, action permissioning, memory, evaluation, and optimization out of the box.
RLM v1 ships with Horizons: open-source Reward Language Models for evaluating agent outputs on unstructured tasks.
MIPRO Offline ships with Horizons: open-source batch optimization of agent prompts and few-shot examples via MIPROv2.
Managed Environment Pools are production-ready with Harbor and Browser pool types, BYOK credential injection, and a full SDK.
Hosted Task Apps let you run evaluation code on Synth infrastructure — no tunnels, no public URLs.
Horizons SDKs published to crates.io, PyPI, and npm.

Horizons: Open-Source Agent Execution Platform

Horizons is a self-hostable platform for building, hosting, and managing persistently running agents. It ships with everything agents need for long-running work:

Context Refresh: Ingest external context into a shared semantic layer on cron or event-driven schedules.
Event Sync: Durable pub/sub with routing, retries, replay, and DLQ to keep agents in sync with external backends and end-user interaction.
Core Agents: Agent registration and execution. Claude Code, OpenCode, and Codex supported day one.
Actions: Propose/approve/deny sensitive agent actions with a full audit trail.
Memory: Voyager-backed long-term memory (graph + vector + summarization) for long-running workloads.

Run locally with docker compose up or deploy to your own cloud. Licensed under FSL-1.1-Apache-2.0 (converts to Apache 2.0 after two years).

Repository: github.com/synth-laboratories/Horizons

Reward Language Models evaluate agent outputs by reasoning about quality in natural language. Describe what good looks like; the RLM evaluates against it. Works for unstructured tasks where traditional metrics break down — no hand-crafted rubrics or regex matching required. Ships with Horizons as the default evaluation system.

Open-Source MIPRO Offline

MIPROv2 runs batch optimization passes that refine agent prompts and few-shot examples. Not a one-time tuning pass — ongoing improvement as business context evolves. Ships with Horizons as the default prompt optimization engine. On the Synth platform, MIPRO Offline extends with ontology-enhanced optimization that incorporates live context from Context Refresh and Memory.

Managed Environment Pools

Environment Pools shipped as a fully managed service. Create a pool, define your task app, and run GEPA or eval rollouts — Synth handles provisioning, credential injection, and artifact persistence.

Harbor Pools: Isolated Docker containers for coding agent evals with 3-second snapshot provisioning.
Browser Pools: Pre-authenticated persistent browser profiles for automation workloads.
Pool Rollouts: Execute GEPA and eval jobs against a pool with create_rollout() — stream events in real-time and collect artifacts.
BYOK Credentials: Store Docker registry tokens, GitHub PATs, and LLM API keys once. Pools handle authenticated pulls and decryption without leaking secrets to sandboxes.

Hosted Task Apps

Your evaluation code can now run directly on Synth infrastructure. Create a hosted task app, reference it by task_app_id in eval and GEPA job configs, and skip the tunnel setup entirely. Task apps support private Git repos and Docker images with credential-based auth.

Horizons SDKs

The Horizons platform SDK is now published across three languages:

Rust: horizons-ai on crates.io
Python: horizons on PyPI
TypeScript: @horizons-ai/sdk on npm

Full coverage for agent management, MCP gateway calls, event pub/sub, context refresh, graph execution, and pipeline triggers.

TL;DR

Horizons launched as an open-source agent execution platform — context refresh, event sync, action permissioning, memory, evaluation, and optimization out of the box.

RLM v1 ships with Horizons: open-source Reward Language Models for evaluating agent outputs on unstructured tasks.

MIPRO Offline ships with Horizons: open-source batch optimization of agent prompts and few-shot examples via MIPROv2.

Managed Environment Pools are production-ready with Harbor and Browser pool types, BYOK credential injection, and a full SDK.

Hosted Task Apps let you run evaluation code on Synth infrastructure — no tunnels, no public URLs.

Horizons SDKs published to crates.io, PyPI, and npm.

Horizons: Open-Source Agent Execution Platform

Horizons is a self-hostable platform for building, hosting, and managing persistently running agents. It ships with everything agents need for long-running work:

Context Refresh: Ingest external context into a shared semantic layer on cron or event-driven schedules.

Event Sync: Durable pub/sub with routing, retries, replay, and DLQ to keep agents in sync with external backends and end-user interaction.

Core Agents: Agent registration and execution. Claude Code, OpenCode, and Codex supported day one.

Actions: Propose/approve/deny sensitive agent actions with a full audit trail.

Memory: Voyager-backed long-term memory (graph + vector + summarization) for long-running workloads.

Run locally with docker compose up or deploy to your own cloud. Licensed under FSL-1.1-Apache-2.0 (converts to Apache 2.0 after two years).

Open-Source RLM v1

Open-Source MIPRO Offline

Managed Environment Pools

Harbor Pools: Isolated Docker containers for coding agent evals with 3-second snapshot provisioning.

Browser Pools: Pre-authenticated persistent browser profiles for automation workloads.

Pool Rollouts: Execute GEPA and eval jobs against a pool with create_rollout() — stream events in real-time and collect artifacts.

BYOK Credentials: Store Docker registry tokens, GitHub PATs, and LLM API keys once. Pools handle authenticated pulls and decryption without leaking secrets to sandboxes.