🛡️ AgentGov — Agent Governance & Observability

A governance control plane for agentic systems. Proxies agent tool calls, enforces policy and budgets, audits every decision, and flags anomalies — so agentic systems become auditable and controllable.

Proof of concept. Synthetic data only. No real systems, no real PII.

The problem

Organizations deploy AI agents faster than they can govern them. Security and compliance teams have no audit trail, no enforcement, and no cost ceiling — so they block production. The bottleneck to enterprise agent adoption is governance, not capability. AgentGov is that governance layer.

What it does

• Enforces — every tool call is allowed, denied, or escalated to a human, per a declarative YAML policy.
• Caps cost — per-agent budgets; an allowed call is still denied if it would breach the ceiling.
• Audits — one immutable, self-explaining row per call. Every decision, every dollar.
• Observes — replay any agent's full action history; flag agents with abnormal denial rates.

Architecture

   Agents / Apps
        |  (tool calls)
        v
   Governance Proxy (FastAPI)
     - enrich (agent -> team)
     - policy engine (allow/deny/escalate)
     - budget gate (cost cap)
     - audit write (always)
        |
        v
   SQLite audit trail  --->  Streamlit dashboard + observability

Layout

agentgov/
├── core/        policy engine, audit core, agent lookup, observability
├── proxy/       FastAPI governance proxy
├── dashboard/   Streamlit audit/anomaly UI
├── synthetic/   seeded synthetic data + request replay
├── policies/    YAML policy rules
├── db/          schema + init
├── governance/  model card, audit schema (trust artifacts)
└── tests/       smoke tests for each component

Run it

# 1. environment
.\.venv\Scripts\Activate.ps1

# 2. build the database + synthetic data
python db\init_db.py --reset
python synthetic\generate.py

# 3. start the proxy (terminal 1)
uvicorn proxy.app:app --reload --port 8000

# 4. drive synthetic traffic through it (terminal 2)
python synthetic\replay_requests.py

# 5. open the dashboard (terminal 2)
streamlit run dashboard\app.py

What it deliberately does not do

See governance/MODEL_CARD.md. In short: AgentGov decides what's permitted, not what's correct; it never auto-approves an escalation; and in this POC all execution is simulated.

Roadmap (beyond POC)

• Real MCP proxy integration (intercept actual tool/MCP traffic)
• Hot-reloadable policy; policy versioning
• Multi-tenant, Postgres-backed, OpenTelemetry traces
• Time-windowed budgets; rate limiting
• Pluggable anomaly detectors with the same explainability contract