openagentscontrol

OpenAgents Control (OAC) is an AI agent framework built on top of OpenCode, an open-source AI coding framework. It is designed around plan-first development workflows: agents analyze a request, propose an implementation plan, wait for human approval, then execute incrementally with validation. The project emphasizes repeatable results by teaching agents a project's own coding patterns before code generation. Core ideas described in the README: - Pattern control: users define coding patterns once (tech stack, API and component examples, naming conventions, standards, security requirements) and agents reuse them, aiming to reduce refactoring of AI-generated code. - Approval gates: agents request approval before writing or editing files, running bash commands, or delegating to subagents. - Editable agents: agent behavior is defined in markdown files that can be edited directly, rather than being baked into proprietary plugins. - Context system: a ContextScout component discovers relevant context files, with a local-first resolution approach. Project-specific patterns live in local context directories, while core standards can fall back to a global install location. - MVI (Minimal Viable Information) principle: context files are kept small (concepts under ~100 lines, guides under ~150, examples under ~80) and loaded lazily to reduce token usage. - Model agnostic: works with various model providers (Claude, GPT, Gemini, MiniMax, local models) and uses the OpenCode default model unless configured otherwise. - Multi-language: agents are described as language-agnostic, adapting to project files; TypeScript/Node.js is primarily tested, with C#/.NET context files also provided, and Python, Go, Rust and others supported but less battle-tested. Included components mentioned: - Main agents: OpenAgent (general tasks and learning), OpenCoder (production development and complex features), SystemBuilder (generates custom AI systems via an interactive wizard). - Specialized subagents that are auto-delegated: ContextScout, TaskManager, CoderAgent, TestEngineer, CodeReviewer, BuildAgent, DocWriter, ExternalScout (fetches live documentation for external libraries), plus category specialists such as frontend, devops, copywriter, technical-writer and data-analyst. - Productivity commands such as /add-context, /commit, /test, /optimize and /context. - A Claude Code plugin (marked BETA) offering a 6-stage workflow, subagents, workflow skills and user commands. Installation is via a shell script (curl piped to bash) with an interactive option, plus an update script. A custom install directory can be specified. The README also documents an example workflow with rough time estimates for discovery, planning, approval, execution and validation, and notes Windows support through Git Bash or WSL. The README includes a comparison table contrasting OAC with Cursor/Copilot, Aider and Oh My OpenCode across features such as pattern learning, approval gates, token efficiency, team standards, agent editability, model choice, execution speed and error recovery. These are the project's own claims and should be treated as such. The project is MIT licensed.