Recover without a redeploy
Failing agents are retried or quarantined automatically, and serious incidents go to your on-call.
For teams running AI agents in production
Govern what your AI agents do.Recover when they don't.
AgentGuardian runs on your infrastructure, outside any one cloud's agent platform. It checks tool calls against your policies before they run and recovers failing agents without a redeploy.
Works with any agent that can make an HTTP call. We reply within 48 hours.
TimeSourceEventStatus
12:04:01finance-agentread_fileALLOW
12:04:02finance-agentexec_shellDENY
12:04:03finance-agentissue_refundAPPROVAL
12:04:05finance-agentread_fileREPEAT
12:04:07detectorloop_detectedINCIDENT
12:04:08engineretryRECOVERING
12:04:35engineagent healthyRESOLVED
Ways to connect
Python today. Any language through the HTTP gateway.
@arep_agentPython decoratorMCPWrapperMCP tool wrapperPolicy gatewayHTTP, any languageREST APIPolicies, incidents, audit
The problem
Finding failures after the fact protects no one.
Agents loop, time out, and drift from what they said they would do. Teams often learn about it from an alert after the damage is done, with no way to stop the next call.
Observability tells you what went wrong. AgentGuardian decides what is allowed to happen, and acts when something breaks.
Why now
Gartner now tracks the category: Guardian Agents.
Gartner published its first Market Guide for Guardian Agents in February 2026, and expects the category to reach 10 to 15 percent of the agentic AI market by 2030. AgentGuardian is a self-hosted guardian layer between your agents and the systems they touch, whichever cloud they run on.
150,000
AI agents Gartner expects per Fortune 500 enterprise by 2028.
13%
Organizations that think they have the right AI agent governance in place, per Gartner.
80%+
Unauthorized agent transactions Gartner expects through 2028 to come from internal policy violations, not malicious attacks.
40%+
Agentic AI projects Gartner expects to be canceled by 2027, citing cost, unclear value, and weak risk controls.
Sources: Gartner on agent sprawl, 2026 · Gartner Market Guide for Guardian Agents, 2026 · Gartner on guardian agents, 2025 · Gartner on project cancellations, 2025
How it works
Connect. Enforce. Recover.
Connect in a few lines
Add one decorator to see every agent run. Put a gateway check in front of the tool calls you want to govern.
agent.py
from arep.sdk import arep_agent
@arep_agent(agent_id="finance-agent")
async def run(task):
# your agent codebefore a governed tool call
POST /gateway/invoke
{ "agent_id": "finance-agent",
"tool": "exec_shell",
"inputs": { ... } }Decide before the tool runs
Each governed call is allowed, denied, or held for human sign-off. Your agent gets the decision before the tool runs.
gateway decisions
read_fileALLOWwithin policyexec_shellDENYblocked by policyissue_refundAPPROVALawaiting sign-offRecover while the agent is running
When a failure is detected, an incident opens and a recovery action starts right away. Severity decides the action.
incident timeline
- 12:04:07P1 OPENEDfailure detected
- 12:04:08RECOVERINGautomatic action
- 12:04:35RESOLVEDagent healthy
Platform
Every agent, one set of rules. And one place to see what broke.
One layer outside your agents for policy, incidents, recovery, and audit, shared by every agent you run.
Enforce before execution
Agents check each governed tool call with the policy gateway and get allow, deny, or wait for approval before it runs.
Incidents, not alert noise
Related failures roll up into one prioritized incident, with the chain of events that caused it.
Catch agents going off-script
Actions that drift from what the agent said it would do are flagged as soon as they happen.
Your fleet, live
Every agent and tool on one live map. Health changes and blocked calls show up within seconds.
Audit-ready by default
A tamper-evident record of every decision, verifiable on demand when compliance asks.
Enterprise governance
Your rules, checked before the tool runs.
Write rules per agent and per tool, and check any call against them without running it. Every decision is recorded for audit.
- Curated guardrail packsPreview
- Ready-made policy packs for the riskiest tool surfaces, so agents are governed from the first call.
- Bring your own policies
- Apply rules per agent, per tool, or across the fleet. Policies are versioned, and a removed policy can be restored with one call.
- Dry-run any call
- Check a sample call against your policies and see the decision without running the tool.
- Humans stay in the loop
- Sensitive actions wait in the approval queue until someone on your team signs off. Slack, PagerDuty, or a webhook tells the right people.
Production guardrails
shell and system access
Data access controls
databases and PII
Deployment approvals
production releases
Payments team
team-defined rules
New policy
Security
Built for your security review.
AgentGuardian runs inside your infrastructure and keeps a verifiable record of every decision it makes.
Runs on your network
Run it with Docker Compose on one server today. A Helm chart and a Kubernetes operator are in development. Agent data stays on your network.
Access you can scope
Agents, operators, and internal services use separate API keys. Each key carries only the permissions its job needs.
Every decision on record
The audit log is hash-chained, so an edit to any past entry shows up when you run an integrity check.
Ready to put a policy check in front of your agents?
FAQ
Common questions.
Which frameworks and languages do you support?
Python agents connect through the @arep_agent decorator, and MCP tools through MCPWrapper. Any language can call the HTTP policy gateway. Integrations for specific agent frameworks are on the roadmap.
Why not use my cloud's built-in agent policy?
If all your agents run on one platform, its built-in controls may be enough. Amazon Bedrock AgentCore, Microsoft Agent 365, and Google's Agent Gateway each tie policy to their own platform's gateway and identity. AgentGuardian runs on your infrastructure and works with any agent that can make an HTTP call, wherever it runs. It also opens incidents and recovers failing agents when a call goes wrong.
Where does AgentGuardian run?
In your infrastructure, on Postgres and Redis. Today it runs on a single server with Docker Compose. A Helm chart and a Kubernetes operator that automates upgrades and scaling are in development. Agent data stays on your network.
What happens when a call needs approval?
The gateway returns a pending decision, the request lands in the approval queue, and your team is notified in Slack. Someone on your team approves or denies it, and the decision is written to the audit log.
Can I see a decision without running the tool?
Yes. A dry run checks a sample call against your active policies and returns the decision without executing anything. Policies are versioned, and a removed policy can be restored with one call.
How does recovery work without a redeploy?
When a detector fires, an incident opens with a severity from P0 to P3. The recovery engine then tells the agent to retry, quarantines the agent, or escalates to your on-call. Retry works for agents that use MCPWrapper with directive polling turned on.
Do you have SOC 2?
Not yet. Because AgentGuardian is self-hosted, you control where agent data lives, and the hash-chained audit log gives your compliance team a record they can verify. Policy packs mapped to NIST AI RMF, the OWASP Top 10 for Agentic Applications, and the EU AI Act, with audit evidence export, are on the roadmap.
Early access
Get runtime control of your AI agents.
We are onboarding teams that run AI agents in production and need a policy layer they can show to compliance.
Received
You're on the list.
We'll reach out within 48 hours to set up your onboarding.
Prefer to talk first?
Book a demo