Verify Five Controls Before Routing Agent Traffic

You are deciding whether agent traffic should run through a vendor control plane. The test is whether the vendor can show the controls, not whether the diagram looks unified. On September 24, 2026, Island announced a $400 million Series F round at a $6.4 billion valuation. Evolution Equity Partners led the round and pushed total investment past $1 billion. The funding shows demand, but it does not answer the operator's question. Palo Alto Networks, along with a growing group of startups, is building products to protect AI systems, agent identities, and the data those systems can reach.

The five-layer story needs proof

Island frames its control plane as a five-part system spanning last-mile control, networking, data, identity, and observability. Each layer is a surface where an agent can fail. The platform uses an endpoint-outward SASE design for zero-trust network access, device posture checks, and inline data loss prevention (DLP), without a traditional VPN. Island's Enterprise AI offering includes identity governance, access controls, guardrails, data boundaries, human sign-off, and cost oversight, along with audit trails across browsers, endpoints, cloud services, and internal resources. The operator still needs to see each control working on the path the agent will use.

A control plane is only as strong as the surface it misses. If an agent can call a cloud service outside the browser, identity and DLP controls may stop applying. If the audit trail cannot leave the vendor, the operator cannot reconstruct an incident for a regulator, customer, or internal review. The path is the proof.

Run the five checks before routing traffic

Use the checklist as a due-diligence pass. Each item is a question for a vendor call, and each answer should be verifiable in a demo or a written response.

• Who issues the agent identity? Request the issuer, scope, and revocation behavior. Identity must map to a human owner and stop working when the owner leaves. Revocation that cannot be shown is a missing control.
• Where does the audit trail live, and can you export it? Confirm the storage location and export path. The trail must cover the agent's actions and decisions. A log that cannot leave the vendor cannot support incident reconstruction after you leave.
• What data boundaries are enforced inline? Identify which DLP rules run on the agent path, not just the endpoint. Request a demo of a policy that stops the agent from reaching protected data. A future roadmap is not a boundary.
• How are human approvals and cost governance enforced? Review the approval workflow and the cost limit that stops work. Approval must require a human, and the cost control must stop the agent before spend exceeds the limit. Advisory controls do not hold in production.
• Does the browser or endpoint choke point match your deployment? Map where the control plane sits in the path. For browser-based agents, the choke point should be there. For agents on endpoints, cloud services, or internal resources, the vendor must show the same controls at those surfaces. A mismatch means the agent can bypass the layer.

Scale is not a control

Island had 1,000 employees and doubled annual recurring revenue (ARR) each fiscal year after launching in 2022. Growth matters when you are choosing a vendor that will still be around after the pilot. It does not replace the checks. A large company can still ship a control plane that cannot revoke an agent identity or export its own audit trail. A vendor that cannot show revocation or export is not ready for production. Treat scale as a durability signal, then test the controls.

Make the pilot prove the path

Send the checklist to the vendor before the pilot. Ask for a written answer for each item, and mark any answer that depends on a future release. The pilot must test the same path the agent will use in production. Keep the answers with the pilot record.