Opencontroller by Lyzr
Opencontroller blog
Answers to the questions enterprises ask about governing AI agents
Agent governance basics
All in this topicWhat should developers learn before building AI agents?
A governance curriculum for AI agent builders: registration, identity, tool scoping, prompt injection, evals and incident duties, with a checklist to sign off.
Who is liable when an AI agent makes a mistake?
When an AI agent gets it wrong, the business deploying it is usually liable to its customers. What the Air Canada chatbot case showed, and the evidence that limits exposure.
How do you report AI agent risk to the board?
A one-page board report for AI agent risk: the metrics to show, trends, incidents, and the decisions you need the board to make.
What is the difference between responsible AI and AI agent governance?
Responsible AI sets principles; AI agent governance turns them into controls on what agents can access and do. A side-by-side comparison of scope, owners and artefacts.
What does a RACI matrix for AI agent governance look like?
A filled-in RACI matrix for AI agent governance: who is responsible, accountable, consulted and informed from registration to retirement.
Discovery and shadow agents
All in this topicHow do you discover AI agents running on-premises?
Find on-premises AI agents from packages, processes, GPU hosts, local model servers and egress logs, with a 10-step inventory flow you can run this week.
How do you discover AI agents and apps through identity provider logs?
Find AI apps and agents in Entra ID, Okta and Google Workspace logs: consent grants, service principals, app role assignments, with a KQL query to start.
How do you find Hugging Face models and tokens used across your organisation?
Find Hugging Face tokens, downloaded models and Spaces: hf_ token scans, Hub cache folders, egress logs, the org token list and audit log export.
Do private endpoints hide AI agent traffic from discovery?
PrivateLink, Azure Private Endpoints and Private Service Connect keep LLM calls off the web proxy. Which signals still reveal AI agents: audit logs, endpoints, DNS.
How do you find and govern voice AI agents in the contact centre?
Voice AI agents hide in contact-centre flows, telephony vendors and outbound dialers. Where to look, what to inventory, and the disclosure and recording controls.
How do you inventory AI agents across subsidiaries with separate cloud tenants?
One AI agent inventory across business units with separate AWS organizations, Entra tenants and Google Cloud orgs: delegated read access, entity tags and a merge plan.
Which shadow AI agents should you deal with first?
Discovery found hundreds of AI assets. A five-factor scoring rubric (data, actions, reach, users, credentials) with a worked example and response times.
How do you find self-hosted LLM servers like Ollama or vLLM on your network?
Ollama listens on 11434, vLLM on 8000, LM Studio on 1234. How to find self-hosted LLM servers with scans, EDR and DNS, and what to check once found.
How do you find AI agents hidden in container images?
Scan registry SBOMs for LLM SDKs, agent frameworks and MCP libraries, check env vars and model weights, then join to running workloads and register owners.
Agent identity and access
All in this topicCan an LLM gateway reuse your existing Okta app for SSO?
One Okta OIDC app can hold several redirect URIs, but a separate app per tool limits blast radius. When to reuse, and the gateway SSO checklist.
What is CIBA and how can it get user approval for AI agent actions?
OpenID CIBA lets an AI agent ask a user to approve an action on their phone. The flow, poll vs ping vs push, binding messages, limits and a checklist.
What are OAuth Rich Authorization Requests and why do they matter for AI agents?
RFC 9396 authorization_details explained for AI agents: a payment example, scopes vs RAR compared, when to use it, and its practical limits.
How should users consent to an AI agent acting on their behalf?
An 8-point consent screen checklist for AI agents, incremental consent, when admin consent must replace user consent, and how to revoke grants.
How should privileged access management work for AI agents?
Vault agent credentials, check them out per task, inject them outside the model, record sessions. A PAM control map for AI agents with a rollout checklist.
What is a joiner-mover-leaver process for AI agents?
Apply identity lifecycle to AI agents: the triggers for joiner, mover and leaver events, the access actions each requires, and a checklist your IAM team can run.
RBAC or ABAC: which access model works for AI agents?
Roles are simple but explode for agents; attributes handle context but need clean data. A side-by-side table, agent-specific attributes and a hybrid pattern.
How does relationship-based authorisation (Zanzibar, OpenFGA) apply to AI agents?
ReBAC answers "can this agent, acting for this user, read this document?" How Zanzibar-style models work for agents and RAG, and check-at-retrieval vs check-at-index.
Runtime control
All in this topicWhat happens when an AI agent invents a tool or its arguments?
Why agents call tools that do not exist or pass wrong arguments, what strict tool schemas guarantee, and the server-side checks and logs you still need.
Which AI agent events should go to your SIEM?
A prioritised list of AI agent events for your SIEM, with the fields to include, volume guidance and example detections your SOC can use.
Should AI guardrails fail open or fail closed?
When a guardrail service times out or goes down, should agent traffic pass or stop? A decision table by risk tier and action type, with timeouts and fallbacks.
How do you monitor hand-offs between agents in a multi-agent system?
Monitor multi-agent hand-offs with one trace across agents, a fixed set of fields per hand-off, and detections for loops, lost context and privilege escalation.
How do you control AI agents that can move money?
Controls for AI agents with payment, refund or transfer tools: amount caps, payee allow-lists, velocity limits, dual approval and reconciliation evidence.
Tools and MCP
All in this topicWhy does an AI agent only see the first page of MCP results?
MCP paginates list operations, not tool results. Why agents treat page one as the whole answer, and how to design and test MCP tools that return lists.
How do you govern long-running MCP tasks?
What the MCP Tasks extension specifies for long-running tool calls—task handles, TTLs, cancellation, auth binding—and the controls enterprises should add.
Should read and write tools live on separate MCP servers?
Trade-offs of splitting read and write MCP tools across servers versus read-only modes, toolsets and annotations, with a decision table for least privilege.
How do you verify an A2A agent card before trusting a remote agent?
A practical checklist for verifying A2A agent cards: well-known location, signed cards (JWS over JCS), declared auth schemes, extended cards and card risks.
What are MCP Apps and what new risks do they add?
MCP Apps let MCP servers return interactive HTML interfaces in sandboxed iframes. What the extension specifies, and the governance questions it raises.
How do you deprecate an MCP server without breaking agents?
A step-by-step way to retire an internal MCP server: find every dependent agent from usage data, set notice periods, use shims, and revoke access last.
Should AI agents be allowed to search and fetch from the web?
How to govern web search and web fetch tools for AI agents: domain allowlists, use limits, URL exfiltration risk, and what to log for each fetch.
Gateways and cost
All in this topicDoes extended thinking work through an LLM gateway?
Claude thinking blocks must pass back unmodified with signatures in tool loops. Where LLM gateways break this, cross-platform rules, cost and a test set.
How do you plan for LLM model retirements in production agents?
Where Anthropic, Azure OpenAI, OpenAI and Vertex AI publish model retirement dates, their notice periods, and a migration checklist with evaluation gates.
How much do LLM-as-judge evaluations cost on production traffic?
A worked cost formula for LLM-as-judge evaluation of agent traffic (traces × sampling × criteria × tokens × price) and the levers that cut it.
Evidence and frameworks
All in this topicBuying and comparison
All in this topicPlatform Guides
All in this topicShould AI agents use Amazon Bedrock API keys or IAM roles?
Use IAM roles for production agents on Bedrock; short-term API keys last up to 12 hours and long-term keys are for exploration only. How to enforce it.
Where can Microsoft Foundry guardrails intervene in an agent run?
Foundry guardrails act at user input, tool call, tool response and output. Which are preview, which tools are covered, latency per point, and the gaps.
How does OpenAI API data residency work for AI agents?
OpenAI API residency is set per project or per request via regional domains like eu.api.openai.com. Which regions process in-region, what is excluded, and a checklist.
How do you enforce Cursor rules across a team?
Cursor Team Rules are set in the dashboard, can be enforced so members cannot turn them off, and win over project and user rules. Limits and a rollout plan.
What does ChatGPT Enterprise Lockdown Mode switch off, and when should you use it?
ChatGPT Lockdown Mode turns off live web, Agent Mode, Deep Research, live connectors and file downloads to cut prompt-injection exfiltration. Who needs it.
How do you govern custom agents in GitHub Copilot?
GitHub Copilot custom agents are Markdown profiles that set tools and MCP servers. Where they live, who can change them and how to review them like code.
How do you add authentication and per-user authorization to LangGraph agents?
LangGraph custom auth uses an Auth object: @auth.authenticate verifies tokens, @auth.on handlers filter threads and assistants by owner. Example and pitfalls.
How do you govern the Amazon Bedrock AgentCore Browser and Code Interpreter tools?
AgentCore Browser and Code Interpreter run in per-session microVMs. The controls to set: network mode, execution role, recording, timeouts and DNS egress.
What do Databricks AI Gateway inference tables log for AI agent audit?
Databricks inference tables log requester, payloads, status and latency per call. Audit gaps: 10 MiB cap, missing 401/403/429/500 rows and delays.
How should AI agents authenticate to HubSpot, and which scopes should they get?
Private app tokens, static auth or OAuth for AI agents on HubSpot: which to use, how to scope by use case, rotation rules and the rate limits that matter.
How do you govern computer use in Microsoft Copilot Studio agents?
Computer use lets Copilot Studio agents click and type on real machines. The controls that matter: machines, access control lists, credentials, reviewers, cost.
Put the framework into practice