Agent Orchestration & Security Template
A reference architecture for running a council of AI coding agents such as Claude, OpenCode and Crush across one codebase with work-queue delegation, automated PR review and hardened CI, alongside research packages for sleeper-agent detection.
This repository is a reference architecture, meant to be studied, forked and adapted rather than contributed to directly, for orchestrating multiple AI agents such as Claude, OpenCode and Crush across a shared codebase with board-driven task delegation, automated pull request review, security hardening and containerized tooling.
It demonstrates measuring and enforcing trust boundaries for autonomous agents through wrapper guards, iteration limits and claim tracking; integrating 19 active MCP servers spanning code quality, content creation, 3D graphics, video editing and speech synthesis; building hardened CI/CD pipelines for agent-authored code with staged lint, test and security checks and multi-architecture Docker builds; and training sleeper-agent model organisms, probing them with residual-stream analysis and linear probes to confirm and localize known-trigger backdoors, though detecting an unknown backdoor is not yet demonstrated.
All code changes in the repository are authored by AI agents under human oversight, and it also covers integrating AI agents into legacy software through runtime injection, shared-memory IPC and overlay rendering. The README recommends reading an included AI safety training guide before working with the autonomous-agent parts of the template.