DEV Community

Dale Chou
Dale Chou

Posted on

Building a governance layer for AI agents

Guardian
Governance infrastructure for autonomous AI agents

AI agents can now:

write code
run shell commands
access databases
call cloud APIs
trigger financial or operational workflows
But most agent systems still look like this:

Agent → Tool → Execution
This architecture is powerful — but dangerously incomplete.

When something goes wrong, most systems cannot answer:

Who approved the action?
What policy allowed it?
Can the decision be replayed?
Is there verifiable evidence?
Guardian introduces a deterministic governance layer between AI agents and execution environments.

The Missing Layer
Modern autonomous systems need more than capability.

They need governance.

Guardian inserts a control plane between agents and execution:

LLM

Agent

Guardian

Execution

Evidence
Every action becomes:

Intent → Policy → Decision → Evidence → Execution
This makes agent behavior:

controllable
auditable
replayable
safe to operate in production
Architecture

Guardian acts as a governance control plane.

Before any action executes:

The agent declares intent
Guardian evaluates policy
A deterministic decision is produced
Evidence is written to a ledger
Only then does execution happen
Core Principles
Guardian is built around five ideas.

Intent — Agents must explicitly declare the action they intend to perform.

Policy as Code — Behavior is controlled by declarative policies rather than hidden logic.

Deterministic Decisions — Guardian returns one of three outcomes: ALLOW, DENY, ESCALATE.

Evidence Ledger — Every decision is recorded as verifiable evidence.

Replay Verification — Decisions can be replayed and validated against policy.

Policy Example
Guardian policies are simple rule declarations.

[
{
"actor": "",
"action": "send_email",
"target": "
",
"effect": "ALLOW"
},
{
"actor": "",
"action": "delete_database",
"target": "
",
"effect": "DENY"
},
{
"actor": "agent_finance",
"action": "transfer_funds",
"target": "*",
"effect": "ESCALATE"
}
]
Why This Matters
AI capabilities are increasing rapidly.

Agents can now:

modify infrastructure
deploy code
move data
trigger financial transactions
Without governance, the risk surface grows dramatically.

Examples of real failure modes:

agent deletes production database
agent deploys unsafe code
agent leaks secrets
agent triggers unintended workflows
Guardian exists to make autonomous systems safer to trust in production.

Quickstart
Run the examples:

python examples/demo.py
python examples/replay_demo.py
python examples/agent_integration_demo.py
Expected behavior:

send_email → ALLOW
delete_database → DENY
transfer_funds → ESCALATE
Example Use Cases
AI Coding Agents — Prevent destructive repository changes or unsafe deployments.

Infrastructure Automation — Control cloud and database operations before execution.

Financial Agents — Require escalation for sensitive actions like fund transfers.

Enterprise AI Workflows — Provide evidence and replayability for AI actions.

Status
Experimental infrastructure project.

Focused on deterministic governance for autonomous systems.

Roadmap
Stage 1 — Core governance engine
Stage 2 — Evidence ledger and replay verification
Stage 3 — Policy DSL and permission model
Stage 4 — Developer integrations
Stage 5 — Hosted governance workflows

Top comments (0)