OpenAgentsControl (OAC) is an open-source AI agent framework built around a plan-first, approval-based development workflow. Instead of letting AI agents generate and apply code changes autonomously, OAC requires agents to propose structured plans that humans review and approve before execution. This design addresses the core reliability problem with AI-assisted development: uncontrolled code generation that drifts from project standards and introduces subtle bugs. ## The Plan-First Approach When a developer issues a coding request through OAC, the system does not immediately generate code. Instead, it activates a multi-stage pipeline: 1. **ContextScout** scans the codebase to discover relevant patterns, coding standards, and architectural conventions. It loads only the information needed for the current task, following the Minimal Viable Information (MVI) principle. 2. **TaskManager** breaks the request into discrete, ordered subtasks with explicit dependencies. 3. **CoderAgent** generates code for each subtask, constrained by the patterns discovered by ContextScout. 4. **TestEngineer** produces test cases for the generated code. 5. **CodeReviewer** validates the output against project standards. 6. **BuildAgent** verifies that the changes compile and pass existing tests. At each stage, the system presents its plan to the developer for approval. No code is written to disk until the human confirms the approach. ## 80% Token Reduction via MVI The Minimal Viable Information principle is OAC's answer to the context window problem. Rather than dumping entire files or repository structures into the prompt, ContextScout extracts only the specific patterns, types, and conventions relevant to the current task. The project reports an 80% reduction in token usage — from roughly 8,000 tokens to 750 tokens for equivalent context — which directly translates to lower API costs and faster response times. ## Multi-Language Support OAC supports TypeScript, Python, Go, and Rust out of the box, with an extensible pattern system for any additional language. Pattern definitions are stored as markdown files in version control alongside the codebase, meaning the entire team shares the same coding standards without manual synchronization. ## Editable Markdown-Based Agents Every agent in OAC is defined as a markdown file. There is no compiled agent logic or proprietary configuration format. Developers can edit agent behavior by modifying markdown documents directly — changing workflows, adding constraints, adjusting communication style, or creating entirely new specialized agents. This approach eliminates vendor lock-in and makes agent behavior fully transparent and auditable. ## Primary and Specialized Agents OAC ships with three primary agents: **OpenAgent**: A general-purpose agent for exploratory tasks and learning new codebases. **OpenCoder**: A production development agent that enforces strict pattern adherence and produces deployment-ready code. **SystemBuilder**: A meta-agent that generates custom AI system configurations tailored to specific project requirements. Beyond these, the framework includes specialized subagents: ContextScout for pattern discovery, TaskManager for workflow decomposition, TestEngineer for automated test generation, CodeReviewer for standards validation, BuildAgent for compilation verification, DocWriter for documentation generation, and ExternalScout for external API and library research. ## Model Agnostic OAC works with any LLM provider: Claude, GPT, Gemini, and local models are all supported. The markdown-based agent definitions are model-independent, so teams can switch providers without rewriting their agent configurations. ## Installation and Profiles A single curl command installs OAC with selectable profiles: Essential (9 components), Developer (19 components), Business (15 components), Full (25 components), and Advanced (32 components). Each profile is tailored to different team sizes and workflow complexity. ## Claude Code Integration OAC is available as a beta Claude Code marketplace plugin, enabling a 6-stage workflow directly within Claude Code sessions. This integration brings plan-first development to teams already using Claude as their primary AI coding tool. ## Community and Adoption With 2,400+ GitHub stars and 210 forks as of March 2026, OAC is gaining traction among teams that need predictable, auditable AI-assisted development. The project is released under the MIT license with 206 commits reflecting active development.