Skip to Content
Developer DocsOverview

Developer Overview

AEGIS (Agent Execution, Graph Intelligence & State) is a stateful runtime platform for enterprise AI agents. It is a domain-neutral compliance runtime: compliance domains are installable packs, entity types are runtime DB-backed definitions, and agent behavior (skills, routers, rules, prompts, settings) is tenant-configurable data. The first installed vertical is oil and gas regulatory compliance in Texas — RRC (Railroad Commission of Texas) spacing, flaring, production, and methane workflows.

What AEGIS Does

AEGIS runs stateful, governed AI agents against a compliance knowledge graph. It provides:

  • AI agent execution with a LangGraph-based stateful pipeline (a single LLM node inside a tool loop)
  • Knowledge graph for entity relationships — per-tenant Apache AGE graphs (wells, leases, operators, permits in the RRC vertical)
  • Memory system with working memory, episodic recall, and an injection ledger for deduplication
  • Human-in-the-loop (HITL) approval checkpoints that fire from skill-attached rules, never optional for regulatory filings
  • Installable compliance packs — each pack declares its skills, checklist templates, and assessment recipes; RRC ships five (Rule 37, Rule 32, Form PR, Flaring Monitor, EPA OOOOb)

Repository Layout

The project is a monorepo with 12 backend services (11 Python + 1 Go, one of which — sandbox-runner — has no TCP port), a Next.js frontend, and shared Python libraries:

aegis/
  agents/                 # 4 legacy YAML agent configs — HISTORICAL, not read at runtime
  data/                   # RRC sample data
  docs/                   # Documentation, specs (docs/specs/), and the Nextra wiki
  frontend/               # Next.js 16 App Router (TypeScript, Tailwind, shadcn/ui)
  infrastructure/         # Docker configs, init SQL, startup/test scripts
  services/
    api-gateway/           # Go reverse proxy (port 8000)
    orchestration-engine/  # LangGraph agent runtime, packs (port 8001)
    memory-service/        # Working + episodic memory, ledger (port 8002)
    knowledge-graph-service/ # Apache AGE graph (port 8003)
    approval-service/      # HITL approvals + audit trail (port 8004)
    ingestion-service/     # RRC scrapers + CSV import (port 8005)
    compliance-monitor/    # Deadline + rule tracking (port 8006)
    flaring-monitor/       # Flaring volumes + emissions (port 8007)
    auth-service/          # Email/password login, JWT, RBAC (port 8009)
    agent-config-service/  # Skills, routers, rules, prompts, settings (port 8010)
    connector-service/     # External data source catalog + credentials + probes (port 8011)
    sandbox-runner/        # Jailed Python execution (unix socket, no TCP port)
  shared/                 # Shared Python models and DB helpers (aegis_shared)
  specs/                  # Phase specifications (R1-R5)

Tech Stack

LayerTechnologyNotes
LanguagesPython 3.12, Go 1.21+, TypeScriptPython for the AI/app services, Go for the gateway
Agent OrchestrationLangGraph 0.1.xStateGraph-based execution pipeline
LLM RoutingLiteLLMBox default anthropic/claude-sonnet-5; model is tenant-configurable via Platform Settings
Web Framework (Python)FastAPIAsync/await, all Python services
Web Framework (Go)net/http stdlibReverse proxy, middleware chain
FrontendNext.js 16, Tailwind CSS, shadcn/uiApp Router, SSE streaming, same-origin /api/v1/*
DatabasePostgreSQL 15 + pgvector + Apache AGESingle DB instance with extensions
Cache / MemoryRedis 7Working memory, injection ledger
Message QueueApache Kafka (Confluent 7.6.0)Async event publishing
Dependency ManagementPoetry (Python), Go modulesPer-service pyproject.toml
Data ModelingPydantic v2Shared models in aegis_shared
Last updated on
Prerequisites