Frequently asked questions
Everything you need to know about Guard, AgentGovern and AI compliance. Can't find your answer? Talk to our team.
Guard
No. Guard records and evaluates by default; nothing is denied until you set a policy to enforce. That's deliberate: a governance tool that blocks on day one gets uninstalled on day two.
The tool call goes ahead. Calls that reached us are recorded and marked so you can find them. If AgentGovern can't be reached, Guard keeps your unsent prompts, tool calls with their inputs and outputs, and Claude's replies on your machine, unmasked, in ~/.agentgovern. It retries them for 48 hours, then keeps any it could not send in ~/.agentgovern/outbox-dead until you delete them or uninstall Guard. By default, Guard never stops your engineers working because our service is unreachable.
Yes, locally. This is a governance tool, not a lock. What you get is a record of what ran and what didn't.
One developer. Guard Free includes one registered developer and 10,000 Guard checks a month, and the account has one member, you. Registering a second developer or inviting a second member asks you to upgrade.
Until the 1st of the next month (UTC), Guard runs in critical-only mode. Your developer still sees critical warnings: those for commands that can destroy data or infrastructure, such as `git reset --hard`, `terraform destroy`, or `DROP TABLE` run from the shell. Other warnings are no longer shown, though an occasional one may still appear during the period. Every warning, shown or not, is still recorded in full. With the current Guard hook, your developer sees a notice in the terminal with an upgrade link, and every critical warning includes the link too.
No. A new account's policies warn and record; none is set to Enforce. If you set a policy to Enforce, it blocks matching tool calls, on every plan, including after the allowance.
Claude Code today. Tell us which assistant your team uses.
AgentGovern
AgentGovern is a compliance-as-code platform that instruments your AI agents (LangChain and LangGraph) and evaluates each action against the EU AI Act policy pack. It generates audit-ready evidence automatically.
The EU AI Act is the world's first comprehensive AI regulation, classifying AI systems by risk level and imposing obligations on providers and deployers of high-risk AI. It applies in stages: prohibited practices since February 2025, general-purpose AI obligations since August 2025, and Article 50 transparency obligations from 2 August 2026. Obligations for stand-alone high-risk systems under Annex III (including many agentic AI applications in financial services) apply from 2 December 2027, deferred from 2 August 2026 by the Digital Omnibus, Regulation (EU) 2026/1744. Annex I product-embedded high-risk systems follow on 2 August 2028.
Yes. AgentGovern provides first-class support for LangChain via a drop-in SDK integration. Two calls, init() and instrument_langchain(), return a callback handler; pass it to your agent through callbacks and its actions are captured, policy-evaluated, and logged for compliance.
AgentGovern supports LangChain and LangGraph via the Python SDK, using instrument_langchain() and instrument_langgraph().
AgentGovern ships a policy pack for the EU AI Act.
Every agent action passing through AgentGovern is logged with a timestamp, the policy evaluation result, the action payload, and the applicable regulatory reference. These logs are aggregated into EU AI Act conformity reports for regulators or internal audit teams.
Behavioral drift occurs when an AI agent's actions change over time in ways that violate compliance policies. For example, an agent begins accessing data it wasn't originally authorized to access. AgentGovern does not detect drift. It records each agent action and its policy evaluation.
Partly. AgentGovern captures your agent's actions, evaluates them against your policies, and records each violation with its regulatory citation for audit review. Each policy can be set to one of four enforcement modes, from least to most restrictive: Log, Warn, Alert and Enforce. Under Log, Warn and Alert, the action goes ahead and the violation is recorded. With the SDK's default settings, Enforce stops an agent only through the EU AI Act pack's two input policies, a prompt that tries to bypass compliance controls and raw PII in a prompt: since SDK 2.6.0, when one of them is set to Enforce and fails on a prompt your LangChain or LangGraph agent sends to its model, the SDK raises PolicyViolation in your code before the model call. With those settings, the pack's other policies don't stop the agent at any mode: the action goes ahead and the violation is recorded. The SDK doesn't stop a tool call from running, and doesn't undo anything that already happened.
The Python SDK is an MIT-licensed SDK on PyPI. https://pypi.org/project/agentgovern/ For the hosted dashboard, talk to us.
Guard has a free plan, Guard Free: one developer, 10,000 Guard checks a month (calendar month, UTC). To go beyond it, upgrade from the dashboard or contact us. For AgentGovern, talk to us. The pricing page at zirahn.com/pricing is the source of truth, not this FAQ.
No. AgentGovern is a hosted service: the SDK runs inside your application and sends agent actions to AgentGovern's hosted policy engine and dashboard. On-premises and VPC deployment are not available.
AgentGovern is built for Financial Services (banking, capital markets, insurance), Healthcare (clinical AI, administrative automation), and any regulated industry deploying agentic AI at scale.