Opencontroller by Lyzr
Opencontroller blog
Answers to the questions enterprises ask about governing AI agents
Agent governance basics
All in this topicWhat risks do AI agents add beyond generative AI chatbots?
How tools, memory, autonomy and delegation create new failure modes for AI agents, mapped to OWASP agentic risk IDs and the control for each.
What is the difference between guardrails, policies and controls for AI agents?
Precise definitions of policies, controls and guardrails for AI agents, one example of each, and where each lives: model call, gateway, platform or process.
What should an internal AI agent governance policy contain?
A section-by-section template for an internal AI agent governance policy, with sample clause wording for scope, roles, risk tiers, runtime controls and retirement.
Discovery and shadow agents
All in this topicHow do you estimate how many AI agents your company has?
Size your AI agent estate source by source: cloud, SaaS platforms, code, API keys, devices and expenses, with a worksheet and confidence ranges.
How do you find AI agents running in serverless functions?
Find agents in AWS Lambda, Azure Functions and Cloud Run functions: model keys in environment variables, SDK packages, model-invoke permissions, egress.
How do you find AI agents hiding in notebooks and data science workspaces?
Find agents built in Jupyter and managed notebooks: scan for LLM SDK imports, scheduled notebook jobs, notebook secrets and saved outputs.
Agent identity and access
All in this topicHow do you limit an AI agent's access to email and calendars?
Scope AI agent mailbox access: delegated vs application permissions, Exchange Online RBAC for Applications, and Google domain-wide delegation risks.
How do you keep customer data separate in a multi-tenant AI agent?
Stop cross-tenant leaks in shared AI agents: propagate tenant context, filter retrieval and memory per tenant, scope credentials and caches, test for leaks.
How should data classification labels control what AI agents can read?
Use sensitivity labels and catalogue tags as agent policy inputs: label-aware retrieval, per-classification model routing, blocking top-tier data externally.
How do you give an AI agent safe read-only database access?
Dedicated role, narrow grants on views, row-level security, statement timeouts and a read replica, with example PostgreSQL grants for AI agents.
Should AI agents use personal access tokens?
Personal access tokens tie an agent to one person's identity and permissions. Compare app installation tokens, bot tokens, workload identity and on-behalf-of access.
Runtime control
All in this topicWhat is the lethal trifecta for AI agents?
The lethal trifecta: private data, untrusted content and external communication in one agent. Worked examples and controls that break each leg.
How do you set confidence thresholds for human review of AI document extraction?
Set review bands from measured accuracy, not raw model confidence. Sample auto-approved items, record each review, and recalibrate after model changes.
How do you stop a model update from silently changing agent behaviour?
Pin model IDs instead of aliases, track provider deprecation dates, and run regression evals before switching so a model update never changes an agent unseen.
What SLOs should you set for production AI agents?
Six SLOs for production AI agents: task success, groundedness, p95 latency, cost per task, escalation rate and policy violations, with how to measure each.
What is a circuit breaker for AI agents?
A circuit breaker automatically stops an AI agent when error rate, spend or tool failures cross a threshold, then tests recovery. States, triggers and thresholds.
How do you stop an AI agent from emailing people outside the company?
Layer a recipient allowlist in the tool, mailbox-scoped send permissions, a mail flow or delivery rule, and alerts. A system prompt alone does not stop external email.
Can you undo what an AI agent did?
Only some agent actions can be undone. Classify tools as reversible, compensable or irreversible, keep an action journal, and control irreversible tools up front.
Tools and MCP
All in this topicHow do you govern MCP resources and prompts, not only tools?
MCP servers expose tools, resources and prompts. Each carries a different risk. A policy per primitive and the inventory fields to track them.
Do MCP roots restrict which files a server can read?
No. MCP roots are informational guidance, not access control, and are deprecated as of 2026-07-28. Restrict file access with sandboxing and OS permissions.
Gateways and cost
All in this topicHow does provider prompt caching cut AI agent costs?
Anthropic and OpenAI price cached prompt tokens differently. Compare write and read multipliers, minimums and TTLs, and how it differs from gateway caching.
When should AI agent workloads use batch APIs?
OpenAI and Anthropic batch APIs cost 50% of standard prices with results within 24 hours. Which agent workloads qualify, which do not, and how to size savings.
How much do MCP tool definitions add to token costs?
MCP tool schemas are sent as input tokens on every model call. How to measure the overhead per agent, a worked example, and ways to cut it.
Should embedding calls go through the LLM gateway?
Embedding calls send whole documents to a provider, spike during re-indexing and are often missing from cost reports. When to gateway them and when a direct path is fine.
Evidence and frameworks
All in this topicHow does the OWASP Top 10 for LLM Applications differ from the Top 10 for Agentic Applications?
OWASP's LLM Top 10 (2025) covers model apps; the Agentic Top 10 covers agents that plan and act. Side-by-side mapping, what is new, and evidence for both.
Does GDPR Article 22 apply to decisions made by AI agents?
GDPR Article 22 covers solely automated decisions with legal or similarly significant effects. When an agent's decision qualifies, exceptions and safeguards.
What SOX IT general controls evidence do AI agents need?
AI agents that touch financial systems fall under SOX ITGCs. A checklist of access, change and operations controls for agents, with the artefact for each.
What does the South Korea AI Basic Act mean for AI agents?
Korea's AI Basic Act took effect on 22 January 2026. What high-impact, generative AI and local representative duties mean for agents, and the evidence to keep.
Platform Guides
All in this topicHow do you log Vertex AI prompts and responses to BigQuery for audit?
Use Vertex AI request-response logging to write Gemini and partner model prompts and responses to BigQuery: setup, sampling, PII and retention.
How does Kubernetes Agent Sandbox isolate AI agents running untrusted code?
Kubernetes Agent Sandbox explained: the Sandbox CRD, templates, claims and warm pools, gVisor and Kata isolation, and the governance it leaves to you.
What data do LangGraph checkpointers store and how long should you keep it?
LangGraph checkpointers save full graph state per thread, often including prompts and tool outputs. What they store, where, and how to set retention.
How do GitHub Copilot enterprise and organization policies interact?
How GitHub Copilot enterprise policies override or delegate to organizations, how conflicts resolve for multi-org users, and an audit checklist.
How do you restrict which MCP servers Claude Code can use?
Claude Code MCP controls for IT: managed-mcp.json for a fixed set, managedMcpServers, and allowedMcpServers/deniedMcpServers with URL and command matching.
How do input and output guardrails work in the OpenAI Agents SDK?
Agents SDK guardrails: tripwires, parallel vs blocking mode, first-agent-only input checks and the tools guardrails don't cover. Production checklist.
Where does OpenAI Agents SDK tracing send your data?
Agents SDK tracing is on by default and sends prompts and tool data to OpenAI. How to disable it, drop sensitive data or export to your own backend.
How do you govern MCP servers added to Microsoft Copilot Studio agents?
Copilot Studio MCP servers are connectors, so Power Platform data policies apply. Auth options, review checklist and inventory steps for admins.
How do you govern MCP servers in the Windows on-device agent registry?
Windows ODR runs packaged MCP servers contained by default. What containment blocks, the gaps (per-host consent, .mcpb, reduced protections) and an IT checklist.
Should AI agents use OpenAI project service accounts?
Production agents should use OpenAI project service accounts, not personal keys that are disabled when staff leave. Setup steps and limits.
What does the OpenAI Audit Logs API record for AI governance?
OpenAI's Audit Logs API records key, project, service account, role and login events, not prompts. Event table, setup steps and detections.
How do you lock down the GitHub MCP server with read-only mode and toolsets?
Lock down the GitHub MCP server with read-only mode, explicit toolsets, lockdown mode and a scoped token. Remote headers and local flags compared.
What does ServiceNow Now Assist Guardian check?
Now Assist Guardian checks offensive content, prompt injection and sensitive topics. Log vs block modes, limits, and what to govern outside ServiceNow.
How do you scope what a Salesforce Agentforce agent can do with topics and actions?
Scope Agentforce agents with narrow topics (subagents), assigned actions, Agent Script available-when gates and a minimal agent user. Checklist inside.
How do you evaluate Databricks agents with MLflow before and after release?
Use MLflow 3 scorers on an evaluation dataset before release, then run the same scorers on sampled production traces. Setup, sampling rates and gaps.
What does Microsoft Foundry Agent Service standard setup keep in your own Azure resources?
Standard setup keeps Foundry agent threads, files and vector stores in your own Cosmos DB, Storage and AI Search. What it covers, needs, and locks in.
What does Amazon Bedrock AgentCore Evaluations measure?
AgentCore Evaluations scores agent sessions, traces and tool calls with built-in and custom judges, online or on demand. What it covers and what it misses.
What changes for governance when you move from Bedrock Agents to AgentCore?
Bedrock Agents is a managed agent configuration; AgentCore is modular infrastructure for any framework. How identity, tools, policy and evidence shift.
How do CrewAI task guardrails work and where do they fall short?
CrewAI task guardrails validate a task's output with a function or an LLM check and retry on failure. Here is how they work and what they cannot govern.
How do you track Claude API spend per agent with Anthropic's usage and cost APIs?
Anthropic's Admin usage and cost endpoints break Claude spend down by API key, workspace and model. Here is how to map that to agents, and where it stops.
Why do Bedrock requests fail through an OpenAI-compatible LLM gateway?
Bedrock calls through a gateway usually fail on auth, model IDs or body format. How Converse, InvokeModel and Bedrock's OpenAI-compatible paths differ.
How do you find every Anthropic API key across workspaces?
Use the Anthropic Admin API to list every API key, its workspace, creator and status, then map keys to owners and agents and disable the ones nobody claims.
Put the framework into practice