AI/ML and Agent Ecosystem

AlphaSwarm separates agent business logic, model governance, evaluation, reinforcement learning, knowledge retrieval, and observability into distinct repositories.

Ownership map

Capability	Owner repos	Notes
Agent runtime/specs	alphaswarm_agents, alphaswarm_core	AgentSpec, AgentRuntime, hash-locked specs, crews, agentic test/debug framework.
Assistants and bots	alphaswarm_assistant, alphaswarm_bots	Custom assistants and bot infrastructure layered on runtime contracts.
MCP integration	alphaswarm_mcp, alphaswarm_ide	Tool exposure and IDE bridge/connectivity for agent workflows.
MLOps/LLMOps/AgentOps governance	alphaswarm_mlops	Registry, promotion gates, eval governance, model/dataset/prompt/run-lineage cards.
Model training/tuning	alphaswarm_models	Custom LLM/model training and tuning library.
Evaluation	alphaswarm_eval, alphaswarm_mlops packages	Continuous-eval engine, scorers, EVAL span telemetry, gate contracts.
Reinforcement learning	alphaswarm_rl	AlphaStream RL project.
Continuous learning	alphaswarm_learning	Scholar/human-agent learning resources; currently has production-path TODO debt.
Data and RAG	alphaswarm_data, alphaswarm_ingest, alphaswarm_kb, alphaswarm_kb_federation, alphaswarm_graph, alphaswarm_index	Market data, catalog, graph sync, RAG ingest, federated retrieval, development assistant index.
Observability	alphaswarm_observe, alphaswarm_observe_js	AGENT/EVAL/RETRIEVAL spans, frontend telemetry, lineage and replay roadmap.

Model training and governance flow

Agent inference and workflow flow

MLOps control-plane posture

alphaswarm_mlops is the governance/control boundary, not a replacement for every runtime package. Its README describes ownership over:

• Model registry, aliases, stages, promotion gates, and MLflow integration.
• LLM gateway seam and provider/catalog governance.
• Token-cost and eval governance.
• Agent run/eval/replay telemetry governance.
• Hash-locked spec registry and advisor gate.
• SR 26-2 model-risk inventory and model/dataset/prompt cards.

Known AI/ML health gaps

• alphaswarm_learning still has P0 Scholar graph/LLM TODO stubs and test-collection failures.
• alphaswarm_kb has a RAG ingest parser provenance/hash drift test issue.
• alphaswarm_kb_federation and alphaswarm_research need additional retrieval/federation/research observability spans.
• alphaswarm_observe has open milestones for evals, lineage, Postgres/read APIs, registries, annotations, alerting, and replay.
• alphaswarm_agents has runtime robustness and PII-redaction debt.

Documentation requirements

1. Publish a canonical AgentSpec lifecycle guide: authoring, hash lock, registry, deployment, telemetry, replay.
2. Publish a model lifecycle guide: dataset card, training run, eval gate, promotion, rollback, drift monitoring.
3. Publish RAG/knowledge lifecycle docs: ingest, parse, provenance hash, graph sync, federation route, retrieval spans.
4. Add traceability tables from model/card/spec IDs to repositories and CI gates.
5. Define minimal production readiness for LLM provider use, local LLM fallback, and cost governance.