Agentic systems
Agent Module v2 (MCP, RAG, Tools, Skills)
File-configured agent platform with RAG, tools, skills, and MCP exposure.
PythonAgentScopePostgreSQLMCPRAGOpenAI-compatible APIs
- Challenge
- Enterprise help agents need RAG over manuals, tool use (web, Jira), and skills — without DB-configured complexity.
- My Role
- Built v2 platform: YAML workspace agents, provider/model abstraction, toolkits, file-based skills with auto-matching, RAG collections per agent, persistent memory, OpenAI-compatible chat API, standalone MCP server, CLI with eval harness flags.
- Architecture
- BaseAgent + AgentFactory + services (Provider, Tool, Skill, RAG) + workspace folders; MCP server exposes web/Jira tools over streamable-HTTP.
- Interesting decisions
- Folders-over-registry (a folder with agent.yaml is an agent); per-agent collection ownership to avoid vector sharing bugs; skills as SKILL.md bundles with companion markdown references.
- Result
- Powers IBPA help, customer-service, QA, and orchestrator agents with shared evaluation pipeline.
- What I Learned
- Convention-based discovery beats registries for small teams; explicit collection keys prevent cross-agent contamination.