Opencontroller by Lyzr
Opencontroller blog
Answers to the questions enterprises ask about governing AI agents
Agent governance basics
All in this topicHow do you measure the ROI of AI agents?
A benefit-minus-full-cost model for AI agent ROI, with a worked example and how to report it by business unit rather than by spend alone.
How do you separate who builds an AI agent from who approves it?
Maker-checker rules for AI agents: who may approve a release, which role conflicts to block, and the evidence auditors expect to see.
Who should own AI agent governance: security, IT or the AI team?
A RACI for 10 AI agent governance activities across the exec sponsor, security, IT, agent owners and risk, with one accountable owner per activity.
What fields belong in an AI agent inventory?
The 13 minimum fields and 12 extended fields for an AI agent inventory, and which ones discovery fills versus which need an owner.
What is AI agent sprawl and how do you contain it?
AI agent sprawl is agents, keys and tools multiplying faster than ownership. Seven indicators to measure it and eight steps to contain it.
Discovery and shadow agents
All in this topicAgent identity and access
All in this topicHow do you link AI usage records to employees when every tool uses a different identifier?
Build one identifier map for AI usage data: which IDs each tool emits, matching rules with confidence levels, and how to handle unmatched records.
How do you run access reviews for AI agents?
A checklist for AI agent access reviews: fields to show reviewers, who signs off, revoke triggers, review frequency and the evidence to keep.
Should AI agents use API keys or OAuth?
API keys vs OAuth for AI agents compared on revocation, scope, audit, user attribution and rotation, with when a key is still acceptable.
How does workload identity federation work for AI agents?
Agents swap a signed OIDC token for short-lived cloud credentials instead of storing keys. How it works on AWS, Google Cloud and Microsoft Entra.
How often should AI agent credentials be rotated?
Default token lifetimes for AWS, Google Cloud, Entra ID, Kubernetes and GitHub, eight triggers for immediate rotation, and how to retire static agent keys.
What does least privilege look like for an AI agent?
A twelve-point least-privilege checklist for AI agents covering identity, tokens, tools, data scopes, thresholds and expiry, with pass conditions.
Are AI agents non-human identities?
AI agents are non-human identities with extra risk. How they compare with service accounts and workload identities, and the six controls agents add.
How does OAuth token exchange work for AI agents?
RFC 8693 token exchange lets an agent act for a user with a narrow, audience-bound token. Request parameters, the act claim, delegation vs impersonation.
Runtime control
All in this topicHow do you govern what AI agents remember?
Treat agent memory as a governed asset: what to inventory, retention limits by memory type, how to honour erasure requests and who may read it.
How do you document which AI traffic paths your gateway actually controls?
Before you enforce AI policy, map every path from agent to model and tool. A traffic-path matrix, an acceptance test per path and common blind spots.
How do you use LLM judges to score production AI agent traffic?
Sampling rates, rubric design and calibration against human labels for LLM judges that score real AI agent conversations in production.
What is a groundedness guardrail and when should it block a response?
How groundedness checks work, how they differ from task adherence, and which action (block, review, warn, log) to take for each type of AI agent answer.
How do you test that your AI agent kill switch actually works?
A 10-step game-day script to test an AI agent kill switch, the five timings to measure, pass criteria and the failures teams usually find.
Tools and MCP
All in this topicHow do you review MCP server updates before agents use them?
Pin versions, hash tool definitions, diff every update and define what forces re-approval. A practical review workflow for local and remote MCP servers.
How do you limit which tools each AI agent can call?
Deny by default, allow-list tools per agent version, constrain arguments and enforce outside the prompt. Worked examples and a policy checklist.
What is an MCP gateway?
An MCP gateway proxies agent tool calls to MCP servers: client auth, per-server credentials, tool filtering, logging and rate limits. How it differs from LLM gateways.
How do you sandbox an MCP server?
Ten steps to sandbox MCP servers: containers, non-root, read-only filesystems, minimal mounts, egress allow-lists, per-server credentials and resource limits.
Is a local stdio MCP server safer than a remote HTTP one?
stdio vs Streamable HTTP MCP servers compared: network exposure, DNS rebinding, process privileges, authentication and a security checklist for each transport.
What should you record for every MCP server in your inventory?
A 15-field MCP server inventory checklist: owner, transport, auth method, pinned version, tool definition hashes, data scopes, consuming agents and review dates.
How do you stop prompt injection that arrives through tool outputs?
Nine defences against indirect prompt injection via tool outputs, ranked by effectiveness, with what each cannot stop and a seven-step rollout.
How do you find local MCP servers on developer laptops?
Where Claude, Cursor, VS Code, Windsurf, Gemini CLI and Codex store MCP configs, and a six-step MDM approach to collect and review local MCP servers.
Should you allow MCP servers from public registries?
Public MCP registries verify publishers, not code safety. A twelve-point approval checklist and a six-step internal allow-list process for enterprises.
What is MCP tool poisoning and how do you prevent it?
MCP tool poisoning hides instructions in tool metadata. The description, rug-pull and shadowing variants, with a control and a ten-point prevention checklist.
How does MCP authorization work?
MCP authorization uses OAuth 2.1, Protected Resource Metadata and resource indicators. The flow step by step, what changed in 2026-07-28, and enterprise implications.
How do you govern agent-to-agent (A2A) calls?
Turn A2A Agent Cards, security schemes and signatures into per-agent call policies: allow-lists, delegated identity, hop limits and audit logs.
Gateways and cost
All in this topicHow do you control which teams can use expensive frontier models?
Tier your models, default every agent to the cheapest tier that passes its evals, give each tier a budget, and grant frontier access by exception.
Should you charge AI costs back to business units or just show them?
Showback vs chargeback for AI spend: when each fits, allocation keys when business-unit tags are missing, and a staged path from one to the other.
How do you attribute GitHub Copilot usage and cost to employees across several GitHub enterprises?
Cover every GitHub enterprise and org, then map GitHub logins to employees using EMU naming rules, SAML identities or verified emails. Gaps included.
How do you find unused GitHub Copilot, ChatGPT Enterprise and Claude seats?
Which activity fields GitHub Copilot, ChatGPT Enterprise and Claude expose for idle seats, their limits, and an inactivity policy you can run monthly.
How do you combine AI seat licences and token spend into one cost view?
Join per-seat AI licences with metered token spend per person, team and agent. A formula, the data each vendor exposes, and a worked example.
How should an LLM gateway handle fallback and retries for AI agents?
Which LLM errors to retry, how to set retry budgets, six equivalence rules for fallback models, and how to stop retries from repeating tool calls.
How much can LLM caching save, and when is it unsafe?
Provider prompt caching vs exact-match vs semantic caching for AI agents: what each saves, a scoped benchmark figure, and seven cases where caching is unsafe.
Buying and comparison
All in this topicHow should an AI governance licence count environments, tenants and VPCs?
A rate-card checklist for AI governance quotes: what counts as a deployment, environment, tenant or VPC, and the services and add-ons to price up front.
Can you govern AI agents in an air-gapped environment?
Yes, for what runs inside the boundary. What works fully offline, what needs a controlled connection, and questions to ask any vendor.
Should you build or buy an AI agent control plane?
A component-by-component build vs buy breakdown for an AI agent control plane, with ongoing maintenance for each part and what to build even if you buy.
Platform Guides
All in this topicHow do you restrict which models teams can use on Vertex AI?
Use the vertexai.allowedModels organisation policy: value format, predict/deploy/tune actions, no wildcards, the 500-value limit, and VPC Service Controls.
How do you restrict which models teams can deploy in Microsoft Foundry?
Two built-in Azure Policy definitions control Foundry model deployments. Parameters, prefix-matching gotchas, model router rules and RBAC.
How do you restrict which Amazon Bedrock models agents can use?
Bedrock retired its Model Access page, so serverless models are open by default. Use an SCP allow-list, and mind inference profiles and Marketplace.
How do you audit Vertex AI and Gemini API calls made by agents?
Vertex AI Data Access audit logs are off by default and never hold prompts. What to enable, what each log records, and how to tie calls to agents.
How do you audit Amazon Bedrock model calls made by AI agents?
Audit Bedrock model calls: CloudTrail vs model invocation logging, what each captures, how to enable logging, endpoint gaps and per-agent token reporting.
How do you replace Azure OpenAI API keys with Entra ID for AI agents?
Move AI agents from Azure OpenAI keys to Entra ID: managed identities, Cognitive Services OpenAI User role, token scope, disableLocalAuth and Azure Policy.
How do you control Claude Desktop extensions and MCP connectors?
Control Claude Desktop extensions and local MCP servers: the organisation allow-list, MDM policy keys like isLocalDevMcpEnabled, precedence rules and gaps.
How do you govern Gemini CLI and Gemini Code Assist?
Govern Gemini CLI with system settings: mcp.allowed, central mcpServers, tools.core, sandbox, enforced auth and OTLP telemetry. Plus Gemini Code Assist admin controls.
How do you govern OpenAI Codex in the enterprise?
Govern OpenAI Codex with managed requirements: sandbox modes, approval policies, network proxy, MCP allow-lists, web search, and cloud, file or MDM delivery.
Put the framework into practice