Protect your organization.
Understand the controls.
Explore what growing agent autonomy means for your organization. Find the questions that matter to business, technology, security and operations leaders, from setting authority to investigating outcomes.
CEO
Decide where agent autonomy creates value and which boundaries the business must own.
Choose another role ↑01Why does our business need an agent control layer?
An agent with tool access can turn a mistaken decision into a real business action. The Agent Control checks explicit permissions before connected tools are called, so your team can expand automation while keeping authority deliberate. It governs the calls routed through it; restricting alternative access is part of the rollout.
02What business outcomes should we measure?
Measure successful governed tasks, approval turnaround, operator effort, policy denials and incidents in the chosen workflow. Compare them with a baseline before expanding access. These are pilot evaluation measures, not a promised return on investment or built-in business analytics.
03What is the right first use case?
Choose a narrow workflow with a clear owner, compatible MCP tools and a measurable outcome. Start with read-only or reversible actions, then introduce consequential actions with human approval. Avoid making the first pilot responsible for an entire business process.
04Who remains accountable when an agent acts?
Your organization owns the permissions, tool access and operating procedures. Administrators configure policy; authorized approvers decide requests requiring human review. The product records decisions and outcomes, but business accountability still needs a named process owner.
05Do we need to replace our AI stack?
No. Your models and agent frameworks remain in place. Compatible clients send tool calls through the Node. Teams must verify authentication, transport and approval handling for each client; integration is not automatically transparent for every agent.
06Where do our data and operational control live?
The Node, operator console, tool credentials and operational records live in infrastructure you manage. Optional remote review shares the selected request context with the Center when configured. Your team should approve that data boundary before enabling review.
07Can we buy and deploy it in production today?
Use the evaluation period to test your intended agents and tools. Review deployment, access controls, recovery and operating requirements with your technical and security teams before committing to a production rollout.
08What should we budget for?
The 30-day trial includes unlimited local operations and 1,000 Center credits per account. A completed standard remote evaluation uses one credit. Include your own infrastructure, integration and operator time when evaluating the total cost of adoption.
09What happens if the system is unavailable?
Calls requiring unavailable authority or remote review must remain blocked rather than gain permission. That can interrupt the governed workflow, so define incident ownership and recovery procedures. The current local runtime is not a qualified high-availability cluster.
10How do we decide whether to expand the rollout?
Require evidence that the pilot enforces permissions, handles approvals, protects sensitive data and recovers without repeating uncertain actions. Expand workflow by workflow only after business and technical owners accept the observed results and operating costs.
CTO
Assess integration, deployment and the limits that matter before adopting a new control point.
Choose another role ↑01Where does The Agent Control fit in our architecture?
Place the customer-operated Node between agents and their MCP tools. The Node authenticates clients, validates calls, checks policy and authorized tool definitions, and dispatches permitted actions. The optional remote review service provides review; the customer portal handles account services separately.
02Which clients and protocols are supported?
The current Node supports MCP tools over stdio and Streamable HTTP. Clients must support the required authentication and, when applicable, asynchronous approval handling. Verify each integration: advanced MCP capabilities and delegated OAuth are not universally supported.
03How do we prevent agents from bypassing it?
Route tool connections through the Node and restrict direct access to upstream tools using network controls, process isolation and credential ownership. The Node cannot enforce policy on a call it never sees. Review alternative execution paths as part of deployment.
04Does adding a tool require custom integration code?
Compatible MCP servers can be configured as upstreams, and their tool definitions are discovered for operator review. Tools outside the supported protocol or clients without the required authentication may need adapters. Validate schemas and transport behavior before enabling them.
05What is enforced locally and what needs remote review?
Local checks cover authenticated client identity, tool permission, definition authorization and any required human approval. Configured remote-review rules send selected context to the Center for evaluation. Local permission checks remain necessary regardless of a remote decision.
06What does our team need to operate?
The persistent installation needs the Node runtime, operator console, PostgreSQL, upstream connections and secure configuration. Your team owns TLS, access restrictions, secrets, backups, monitoring and recovery. The installation guide describes local evaluation; production acceptance needs a separate operating review.
07Can we scale it as a highly available cluster?
The current persistent runtime uses a single-runtime database lock; a multi-worker high-availability cluster is not implemented or qualified. Size and benchmark the intended workload, and treat availability requirements as a pilot acceptance criterion rather than assuming horizontal scaling.
08What latency will it add?
Every governed call adds authentication, policy and catalog checks; remote review adds network and evaluation time, and human approval adds operator wait time. There is no published latency SLA. Benchmark representative workloads and distinguish automated checks from approval wait time.
09What execution guarantees can we rely on?
Approval continuation reuses the stored action and prevents concurrent continuations from creating another dispatch. Uncertain outcomes are retained without automatic resend. The final local check and an external effect are not one transaction, so external exactly-once execution is not guaranteed.
10How should we qualify a rollout?
Test real clients and tools, denied calls, changed definitions, approval resumption, review outages and backup restoration. Establish throughput and recovery targets. Accept the pilot against those results before deploying broader or more consequential workloads.
AI Engineer
Connect agents correctly and make denial, approval and uncertainty explicit in their execution flow.
Choose another role ↑01How do I connect my agent?
Connect the agent to the Node instead of directly to each upstream MCP server. Use the configured client credential and the supported transport. The installation guide covers the persistent local runtime; confirm that your client can send authentication and preserve approval state.
02How is my agent identity established?
For persistent MCP HTTP sessions, a bearer credential selects the configured client identity. Stdio calls use the supported token metadata extension. Agent-supplied identity claims do not grant authority. Protect the credential: a process holding it can act as that client.
03Why is a discovered tool missing or denied?
Discovery does not authorize use. The exact tool definition must be reviewed and an applicable policy must allow the authenticated client to call it. A matching denial takes precedence. Check the catalog, client identity and policy before changing the agent prompt.
04How should my agent handle human approval?
Send a stable requestId for an action requiring approval. A pending response means the action has not executed. Preserve the returned approvalId, query acl__approval_status and use acl__approval_resume after approval. Resume uses the stored action; it does not accept replacement arguments.
05Can I retry a request safely?
For an approval request, reuse the same requestId with the same action and identity to retrieve its existing state or retained result. Different content conflicts. Ordinary calls do not have universal deduplication, and uncertain external effects must not be retried automatically.
06What happens when a tool definition changes?
A new or changed definition loses authorization until an operator reviews it. The Node refreshes the catalog and checks definitions before dispatch. This does not detect an upstream implementation change that leaves the reported definition and version identical.
07Will the control layer change my model or prompts?
It does not replace your model or agent framework. Your execution loop does need to handle authenticated calls, denials, pending approvals and uncertain outcomes. Treat these as explicit states rather than asking the model to work around a blocked action.
08How are tool arguments checked?
Inputs are validated against the supported tool schema and are bound to the authorized action. Basic local permissions identify clients and tools; they are not universal business validation of every argument. Use explicit review criteria and tool-side constraints for business-specific requirements.
09What context goes to remote review?
When configured rules require review, the Center receives the selected request context and operator-defined criteria or evidence. Local action binding and remote context serve different purposes. Check redaction and required evidence with the operator; do not include credentials or unnecessary sensitive data.
10How should I test the agent integration?
Exercise an allowed call, a denied call, a definition change, pending and rejected approvals, resumption and a timeout with an unknown outcome. Verify actual upstream effects. The agent should preserve identifiers and report blocked or uncertain states without switching to an ungoverned tool path.
CISO
Review identity, privileged access, data sharing and the evidence behind enforcement.
Choose another role ↑01What security boundary does this enforce?
The Node governs authenticated calls to connected tools before dispatch. It requires explicit permission and authorized definitions. It does not secure bypass paths, the model itself or a compromised upstream system. Network restrictions and least-privilege tool credentials remain part of the security design.
02How are client permissions enforced?
Each authenticated client is mapped to configured identity and evaluated against local policy. Without a matching allow, the call is denied; a matching deny takes precedence. Definition authorization is a separate requirement, so a broad policy does not authorize an unreviewed tool.
03Where are secrets and sensitive data stored?
Tool credentials and operational records belong to the customer-managed installation. The persistent runtime encrypts stored action payloads and retained results with the installation key. Protect the runtime, database, backups and key together; encryption does not remove administrator trust.
04What data leaves our environment?
Optional remote review sends the selected request context and operator-defined review information to the Center. The account portal is a separate surface and does not receive operational data by default. Review payload selection, redaction, retention and provider arrangements before enabling remote review.
05Can prompt injection override a denial?
Agent instructions cannot grant local policy authority or approve a tool definition. However, an injected prompt can still cause misuse within permissions already granted. Narrow permissions, argument constraints in tools and required review reduce exposure; this is not a promise to eliminate prompt injection.
06How are human privileges separated?
The persistent console has Administrator, Approver and Auditor roles. Administrators configure the installation, approvers decide requests and auditors have read-only redacted access. MFA, enterprise SSO and per-tenant operator role separation are not implemented in the current local unit.
07What happens when authority is revoked?
Changes to operators, credentials or policy affect later authorization checks and can invalidate outstanding approvals. A revocation cannot undo an external action that has already crossed the execution boundary. Test revocation behavior and define upstream incident response for completed effects.
08Can the remote review service execute our tools?
The Service returns a review decision; the customer-operated Node owns tool dispatch and enforces the outcome. Tool credentials remain under customer control. A remote allowance does not replace local authorization, and required review unavailability does not grant permission.
09What audit evidence is available?
The local installation records policy decisions, approval actors and execution outcomes with redacted audit access for auditors. Establish retention, access and backup procedures with the installation owner. Do not assume a certified tamper-proof archive or a turnkey SIEM integration.
10Does this satisfy our compliance requirements?
It can contribute controls and evidence, but no certification or automatic compliance outcome is claimed. Assess it against your own requirements, including identity, retention, supplier review and recovery. Product decisions do not replace your security assessment or legal review.
COO
Design approval workflows and recovery procedures that keep teams accountable as automation grows.
Choose another role ↑01How do we keep approvals from becoming a bottleneck?
Choose a narrow workflow and classify actions by consequence. Permit routine actions only within explicit permissions, and require approval where a person must judge the effect. Measure queue age and reviewer effort during the pilot before increasing volume; automatic staffing or routing is not promised.
02Who should configure, approve and review activity?
Assign an administrator to maintain policy, approvers to judge consequential requests and auditors to review redacted activity. Also name a business process owner and an incident owner. Product roles provide access boundaries; your operating procedure defines responsibilities and coverage.
03What does an approver actually authorize?
The approver reviews a specific stored action, its parameters and the reason for the decision. Resumption uses that exact action and rechecks authority. Changes to action content or relevant policy require a new valid authorization rather than reusing an old approval.
04What should we do when an action is blocked?
Check the reason: client identity, missing permission, unreviewed definition, required review or pending approval. Involve the responsible operator and process owner. Resolve the underlying condition through the controlled workflow instead of handing the agent direct upstream access.
05How do we handle an unknown execution outcome?
Pause the workflow and inspect the upstream system and local audit evidence. The action may have happened even if its response was lost. Record what is known and follow an agreed reconciliation procedure; do not treat a timeout as permission to resend.
06What happens during a Center outage?
Actions requiring remote review stay blocked or pending until a valid outcome is available. Actions eligible for local authorization may continue if local authority remains healthy. Define which workflows depend on remote review and how teams communicate delays.
07What must be backed up and restored?
Preserve the database, secure configuration and installation encryption key using protected backups. Use the documented restore procedure, which revokes sessions and preserves uncertainty for actions interrupted during execution. Test restoration in an isolated environment before relying on it operationally.
08How should we introduce policy changes?
Have the process owner and administrator review the affected clients, tools and approval requirements. Test the intended allow and deny behavior before rollout, and inform approvers. Changes can invalidate outstanding approvals, so plan how pending work will be revisited.
09Which operational indicators matter most?
Track completed tasks, pending approval age, denials, review failures, unknown outcomes and recovery time. Compare workload with available approver capacity. These are recommended operating measures; the product does not currently promise a complete operations analytics dashboard.
10When is a workflow ready for wider use?
Require documented ownership, tested permissions, approver coverage, incident procedures and a successful backup restore. Confirm how the agent behaves when blocked or uncertain. Expand only after the team can demonstrate safe execution and manageable workload with real tools.
Ready to evaluate a workflow?
Start with the architecture, then follow the local installation guide.
Explore Node & Center →Open installation guide →