Opencontroller by Lyzr
Opencontroller blog
Answers to the questions enterprises ask about governing AI agents
Agent governance basics
All in this topicWhen should an AI agent be given more autonomy?
Promotion criteria for AI agent autonomy: eval pass rate, override rate, incidents, time at level, live scores, plus rollback triggers.
How do you govern AI agents inherited in a merger or acquisition?
A day 1, day 30 and day 100 checklist for AI agents inherited in an acquisition: discovery, credential rotation, owners, policy and decommissioning.
What is the difference between an AI agent, an AI assistant and a copilot?
AI agents, assistants and copilots differ in who decides the next action and what they can touch. A comparison table and the governance each needs.
How do you run an AI agent governance committee?
Membership, decision rights, a 60-minute standing agenda and the monthly metrics pack for an AI agent governance committee, plus what to delegate.
What should an AI agent's scope statement say?
A field-by-field template for an AI agent scope statement: purpose, allowed and prohibited actions, data, tools, autonomy, escalation, owner and review.
Discovery and shadow agents
All in this topicWhy does an AI inventory show models an account cannot actually call?
Listed, callable and invoked are three different model states on AWS, Azure and Google Cloud. How to record each so your AI inventory is accurate.
How do you find MCP servers exposed to the internet?
How to scan your attack surface for MCP servers that answer without authentication, what causes exposure, and the five steps to take when you find one.
How do you find custom and fine-tuned models in your cloud accounts?
The list APIs for custom models, tuning jobs and endpoints on AWS, Azure and Google Cloud, how to cover every account, and what to record for each model.
How do you find shadow AI subscriptions in expense reports?
Find AI tools bought on corporate cards: keywords and merchant codes to search, how to confirm recurring subscriptions, and how to log them in your inventory.
Agent identity and access
All in this topicCan AI agents use multi-factor authentication?
Why interactive MFA does not fit AI agents, the workload equivalents that achieve the same goal, and how to word the MFA policy exception for agents.
How do audit logs show which user an AI agent acted for?
Record both the user and the agent on every action: the RFC 8693 act claim, propagating user context through tool calls, and the log fields to capture.
Should AI agents use sender-constrained tokens like DPoP?
Bearer vs DPoP (RFC 9449) vs mTLS-bound tokens (RFC 8705) for AI agents: how each works, where each fits agent runtimes, and what binding cannot solve.
How do you stop an AI agent token from being used against the wrong API?
Bind agent tokens to one API with RFC 8707 resource indicators and audience checks, why MCP forbids token passthrough, and a test checklist per API.
Runtime control
All in this topicHow do you regression-test an AI agent after changing its prompt?
Catch prompt-change regressions with a golden dataset, deterministic checks, calibrated LLM judges, fixed pass thresholds and CI release gates.
What can you do to a misbehaving AI agent short of killing it?
A six-step response ladder for misbehaving AI agents, from alert to kill, with example triggers and who can authorise each step.
Do prompt injection detectors stop indirect prompt injection?
Detectors reduce indirect prompt injection but adaptive attacks get past them. Learn the design patterns that contain injection and how to test both.
How do you stop an AI agent from deleting production data?
Seven layers that stop AI agents deleting production data: separate credentials, database privileges, deny rules, soft delete, backups and confirmation steps.
Tools and MCP
All in this topicCan you manage one MCP server configuration across Copilot, Claude Code and Cursor?
Each coding agent reads its own MCP config file. How to run one approved MCP catalogue across GitHub Copilot, Claude Code and Cursor without drift.
How do you check open-weight models before using them in agents?
A pre-use checklist for open-weight models: safe file formats, hub scan limits, remote code, licence terms, hash pinning and behavioural evaluation.
How do you check agent skills for malicious instructions before use?
A 14-point review checklist for SKILL.md files and bundled scripts: hidden instructions, setup steps, network calls, credential access and provenance.
What is the difference between agent skills and MCP servers?
Skills package instructions and scripts an agent loads on demand; MCP servers give it live tool connections. Risk profiles and a governance table for both.
What is the security checklist for hosting a remote MCP server?
A 14-point checklist for a remote MCP server: OAuth per the MCP spec, audience validation, no token passthrough, Origin checks, rate limits and logging.
What happens when two MCP servers expose tools with the same name?
MCP tool names are only unique per server. How collisions and cross-server tool shadowing happen, how clients namespace tools, and registry rules to stop it.
How many tools should one AI agent have?
Vendor guidance says fewer than 20 tools up front, with accuracy dropping past 30–50. How to cut the list with tool search, per-task tool sets and sub-agents.
Should agents call internal APIs through MCP or OpenAPI tools?
MCP vs OpenAPI tools for internal APIs: how they compare on discovery, auth, schema reuse, governance hooks and client support, plus a rule for choosing.
How do you govern AI agents that control a computer or browser?
Run computer-use agents in isolated low-privilege VMs, allowlist domains, keep credentials out, retain screenshots deliberately and confirm high-impact actions.
How do you govern code execution tools given to AI agents?
Sandbox rules for agent code tools: no network by default, CPU, memory and time caps, ephemeral files, hosted vs self-hosted interpreters, and what to log.
What should you log for every MCP tool call?
A 14-field checklist for MCP tool call logs mapped to OpenTelemetry MCP and GenAI conventions, plus redaction rules for arguments and results.
Gateways and cost
All in this topicHow do you tie every AI agent to a cost centre when it is approved?
The seven approval fields that make AI agent chargeback automatic, how the cost centre flows to keys, gateway routes and cloud tags, and a backfill plan.
Why do long AI agent sessions get so expensive and how do you cut the cost?
Agents resend the whole context every turn, so session cost grows roughly quadratically. A worked example and seven levers to cut it.
Buying and comparison
All in this topicShould you use open-source or commercial tools for AI agent governance?
A category-level comparison of open-source and commercial AI agent governance tools on coverage, integration effort, support and evidence, plus hybrid patterns.
What data processing terms should you require from an AI governance vendor?
A clause checklist for DPAs with AI governance vendors that see prompts and traces: data categories, retention, subprocessors, residency, no training, deletion.
Do you need an AI agent control plane if you only use one agent platform?
Where a single agent platform's native controls are enough, where they stop (models, keys, SaaS AI, coding agents) and five criteria for deciding.
What are red flags when buying AI agent governance software?
Twelve red flags in AI agent governance software, from questionnaire-only discovery and no enforcement path to "compliant" claims, with what to ask instead.
What should you ask to see in an AI agent governance platform demo?
Ten live tasks to request in an AI agent governance demo, from discovery in your own account to blocking an agent and exporting evidence, with a scoring sheet.
Which budget should pay for AI agent governance?
Security, AI platform, shared or chargeback: the options for funding AI agent governance, the case for each, and how spend attribution changes the maths.
What are the categories of AI governance tools?
A map of AI governance tool categories: AI-SPM, LLM gateways, GRC modules, model risk tools, AI DLP and browser security, and agent control planes.
Who should be on the team that evaluates an AI governance platform?
The roles to include when evaluating an AI governance platform, what each should score, and a weighted decision matrix you can adapt.
What SLA should you ask for from an AI governance platform?
SLA checklist for AI governance platforms in the request path: availability, support response by severity, RPO/RTO, fail-open or fail-closed, and credits.
Platform Guides
All in this topicHow do you track Cursor usage and spend across a team?
Use the Cursor Admin API: members, daily usage, spend and usage events endpoints, their limits, and how to join them to cost centres.
How do Databricks Apps service principals work for AI agents?
Each Databricks App gets its own service principal. How app vs user authorization works for agents, and how to scope Unity Catalog grants.
How do you govern the Snowflake-managed MCP server?
Govern Snowflake's managed MCP server with OAuth, per-tool grants and least-privilege roles. What it exposes, its limits, and the checks to run.
Can you run Claude Code through Amazon Bedrock or Google Vertex AI?
Yes. Set CLAUDE_CODE_USE_BEDROCK or CLAUDE_CODE_USE_VERTEX, grant the right IAM, and pin model versions so coding agent traffic stays in your cloud account.
How do you find unused permissions on AI agent IAM roles in AWS?
Use IAM Access Analyzer unused access findings to spot permissions your AgentCore and Bedrock agent roles never use, then trim them safely.
Does Amazon Bedrock cross-region inference move data out of your region?
Geographic vs global Bedrock inference profiles, what AWS says about where prompts are processed and stored, and the IAM and SCP controls to enforce residency.
What are Amazon Bedrock Automated Reasoning checks and when do they help agents?
How Bedrock Guardrails Automated Reasoning checks verify agent answers against policy rules, what valid, invalid and satisfiable mean, and their limits.
Should AI agents use Azure OpenAI provisioned throughput or standard deployments?
When AI agents should use Azure provisioned throughput (PTU) vs standard deployments, how spillover works, and how to attribute shared PTU cost per agent.
Can Apigee be used as an LLM gateway for AI agents?
Apigee's LLMTokenQuota, PromptTokenLimit and Model Armor policies, the prerequisites, and what an API gateway still cannot govern for AI agents.
What GitHub Enterprise permissions does an AI governance tool need?
The read-only GitHub App permissions and classic token scopes an AI governance connector needs for Copilot metrics, seats and the enterprise audit log.
Why does Amazon Bedrock return AccessDenied or ValidationException through an LLM gateway?
Map Bedrock AccessDeniedException, ValidationException and expired-token errors behind an LLM gateway to their causes, with a 7-step triage checklist.
Put the framework into practice