Resources / Agentic AI / Architecture

Architecture

The Agentic AI Platform layers a graph workflow runner on top of the ReAct agent, with packs configured in Agentic Studio and stored in SQLite.

Components

Studio UI ──► /studio/api (models, skills, agents, packs)
Workspace ──► GET /agent/stream
Demo      ──► POST /agent/demo
                    │
                    ▼
             AgentService (ReAct)
                    │
       ┌────────────┼────────────┐
       ▼            ▼            ▼
 GraphRunner    ModelRouter   ToolExecutor
       │            │
       ▼            ▼
 PackRegistry   Bound LLM
 (SQLite packs) (OpenAI / Anthropic / custom)

Data model

Table Purpose
agent_models Provider bindings (openai / anthropic / custom)
agent_skills Prompt + tool allowlist packs
agent_defs Named agents (model + skills + tools + prompt)
workflow_packs Pack metadata (pdlc / sdlc / custom)
workflow_nodes Graph nodes: start, agent, gate, branch, end
workflow_edges Directed edges with optional conditions
project_workflow_state Per-project pack_id, node, completed stages

System packs (PDLC, Cyber Security, Product Research) are seeded on init_db and are cloneable, not edited in place.

Runtime

  1. PackRegistry — load system + user packs
  2. GraphRunner (app/agent/graph_runner.py) — resolve current node; handle gates/branches; advance on request
  3. SkillComposer — merge skill prompts and intersect tool allowlists
  4. ModelRouter — bind the agent’s model to an LLM provider
  5. AgentService — ReAct loop with filtered tools; SSE events include pack, node, agent, workflow, awaiting_gate

Extensibility

Open Agentic Studio to:

  1. Add models, skills, and agents
  2. Clone PDLC or create a blank / SDLC pack
  3. Edit the SVG graph (agent / gate / branch nodes)
  4. Export/import JSON packs
  5. Run to bind a project and open the workspace

No Python registry edits are required for user packs.