What I Built

Most developer setups force engineers to choose a single coding assistant harness or manually copy-paste context between different agentic environments.

Multiharness solves this fragmentation. It is an open-source multi-agent orchestration framework written in Go designed to delegate tasks, manage iterative review loops, and coordinate coding sessions seamlessly across environments like Codex, OpenCode, and local LLM backends.

The Problem: Single-agent harnesses often get stuck in local optima, struggle with verification loops without consuming massive context windows, and create lock-in around closed ecosystems.

Who It’s For & Why They'd Pay: Engineering teams, technical founders, and freelance developers looking to automate complex refactoring, test-generation pipelines, and cross-model code review loops without being tied to a single proprietary vendor or incurring massive API markups.

Code

Multiharness is fully open-source. Explore the architecture, inspect agent harnesses, or contribute:

ErOr-0 / Multiharness-Core

Multiharness-Core is a central orchestration framework driven by an agentic workflow.Codex acts as the lead Planner to design architecture and delegate tasks, while Opencode serves as the Executor to implement the codebase according to Codex's blueprints.

Multiharness Core

magent coordinates coding tasks across Codex, Claude Code and Muse Code. Choose a folder and one agent. In direct mode (the default), Multiharness passes your task to that CLI, lets it handle planning/editing/testing, and shows its final response. Follow-ups reuse the same agent session; /new starts fresh.

Use /set mode team (or --mode team) when you want separate planning implementation, deterministic validation, review and bounded repair. The team workflow also owns folder snapshots and independent change evidence.

Both modes support folders without Git.

Muse Code (Meta subscription)

Select muse independently for planner, implementer or reviewer in /config. Each role has its own model and reasoning setting. For an all-Muse team:

/set mode team
/set planner-harness muse
/set planner-reasoning low
/set implementer-harness muse
/set implementer-reasoning medium
/set reviewer-harness muse
/set reviewer-reasoning high
/login muse

The default model is muse-spark-1.3. Use each role's *-model setting to…

View on GitHub

How I Built It

Multiharness is built with performance and composability at its core:

Engineered in Go: Selected for low-latency concurrency, deterministic process management, and native cross-platform binaries with zero external runtime dependencies.

Open-Source AI & Local Inference: Multiharness supports self-hosted open-weight models (via vLLM and Ollama) alongside open harnesses like OpenCode. It routes architecture, implementation, and test-review roles across specialized models depending on context window and reasoning requirements.

Delegation & Review Loops: Uses structured task state machines where lead agents break down feature requests into atomic diffs, dispatch them to worker agents, and trigger automatic linting and test validation stages before merge proposals.

Why Does Open Innovation Matter?

Open innovation is critical when building developer tooling:

Avoiding Vendor Lock-In: Developers should not be trapped within a single closed harness API that changes pricing, rate limits, or deprecates models overnight.

Context Privacy & Local Execution: Engineering teams often handle sensitive codebases that cannot leave internal infrastructure. Open-weight models and open agent harnesses allow Multiharness to execute entirely on-prem or on developer workstations.

True Modularity: Open architectures let developers plug in customized prompt templates, custom evaluation harnesses, and distinct model weights tailored for specific languages and domain logic.

DEV Community

Multiharness: Orchestrating Autonomous Multi-Agent Workflows Across Open-Source Coding Harnesses