OpenCode Harness: Capability-Adaptive Architecture for Multi-Agent Collaborative Programming

OpenCode Harness is a multi-agent collaborative programming framework that automatically discovers project-level capabilities via an orchestrator, supports 6 professional agent roles and a skill system, and enables adaptive AI-assisted programming workflows. Its design principles include role division, capability adaptation, skill reuse, and memory accumulation, providing a flexible and efficient collaboration solution for complex development tasks.

Early AI programming assistants (e.g., GitHub Copilot) adopted a single-agent model, only providing code completion with limited capabilities. With the development of large models, multi-agent architectures have emerged, decomposing complex tasks into specialized agents for collaborative completion. OpenCode Harness is a concrete implementation of this architecture, defining 6 agent roles and supporting collaboration.

OpenCode Harness defines 6 specialized agents:

• Assistant (Main force, Deepseek V4 Flash, multi-file analysis/refactoring)
• Architect (Architecture design, Deepseek V4 Pro, scheme comparison/risk assessment)
• Reviewer (Code review, Deepseek V4 Pro, bug/security vulnerability detection)
• Scout (Lightweight search, MiniMax M2.5 Free, file lookup)
• Worker (Simple modifications, Big Pickle Free, small-scale adjustments)
• Orchestrator (Core, task decomposition and agent scheduling) Role division optimizes cost and efficiency: lightweight models are used for simple tasks, while powerful models for complex ones.

The system is divided into a core layer (provided by Harness: general agent definitions, skills, orchestration logic) and a project layer (provided by specific projects: configuration files AGENTS.md/CONTEXT.md, exclusive skills, knowledge base). When the orchestrator starts, it automatically scans the project layer to discover capabilities; if none exist, it gracefully degrades, flexibly supporting rapid startup of new projects and capability expansion of mature projects.

Skills are reusable capability units that encapsulate specific functions. Common skills include:

• grill-task: Alignment interrogation before coding to reduce rework
• html-diagram-gen: Convert architecture to HTML diagrams
• mermaid-ascii-diagram: Generate ASCII flowcharts The project layer can define exclusive skills in the skills/ directory, which the orchestrator discovers and loads via AGENTS.md, supporting knowledge precipitation and cross-project sharing.

The built-in memory system stores project knowledge (coding standards, design patterns, pitfalls, etc.) in Topic files under the memories/ directory. The orchestrator queries/records memories to ensure agents share consistent knowledge, simulating the way human teams accumulate and share knowledge, and timely reminding users to submit new memories.

The orchestrator adopts a radical decomposition strategy, splitting complex tasks into independent subtasks and delegating them to different agents for parallel execution (e.g., assigning front-end and back-end tasks simultaneously), breaking the linear time growth of serial execution and improving collaboration efficiency.

Installation is simple: npx skills@latest add moraluco/opencode-harness, select orchestrator as the main agent, and create configuration files according to conventions. The project is open-source under MIT license, providing a reference for AI programming tools. In the future, multi-agent architecture may become mainstream, and this project lays the foundation for community exploration.