agenticswarm-brain

Portable, repo-aware orchestration scaffold for building a self-maintaining multi-agent brain inside any codebase.

AgenticSwarm Brain is designed to be dropped into a repository, inspect that repository plus the globally installed agent/skill/MCP/CLI/model surface, populate a portable capability graph, and keep itself updated through automated startup rechecks and maintenance.

• Audits any repository it is installed into
• Discovers local and global agents, skills, plugins, MCPs, CLIs, and model surfaces
• Builds portable registries the orchestrator can route against
• Validates schemas and cross-file contract consistency
• Runs maintenance automatically: memory reconciliation, topology pruning, and distillation prep
• Installs background startup rechecks on supported platforms so new installs are detected without manual intervention

Most agent frameworks stop at prompts, tools, or task routing. This project is meant to operate more like a reusable orchestration layer:

• portable across repos
• aware of its environment
• able to prefer the cheapest safe execution path
• capable of learning from route history
• structured enough to validate, version, and automate

The goal is not just “run agents.” The goal is to give a repository a persistent, inspectable, self-updating brain.

The brain inspects:

• repository-local agent definitions
• repository-local skills and scripts
• plugin manifests
• MCP configs and MCP executables
• installed CLI tools
• available model runtimes and cached model catalogs

The system is built around:

• semantic routing
• draft-and-verify loops
• worker / critic / expert escalation
• dynamic DAG generation
• historical route replay
• context compression

After bootstrap, the brain can:

• recheck watched install roots on startup
• detect newly installed capabilities
• refresh capabili