2026-11-13 · ← News
GPT-6 Astra lets agents into legacy IT portals while technicians keep control
Shield is using GPT-6 Astra for thousands of monthly runs inside IT portals without APIs and reports execution 35 to 40% faster than with earlier models. The important part is the delegation boundary: technicians decide what the agent may finish and where it must wait for a person.
The image could not be loaded.
Shield is using GPT-6 Astra in its Forge platform to operate older web portals that do not expose APIs. The agent reads the screen, clicks and types much like a technician. The company reports thousands of these runs each month across its IT service providers.
The agent navigates the portal, but technicians set the rules
Forge assigns supported tasks to a computer-use agent. Technicians define in advance which steps may run autonomously and which require review or approval. Phone calls, on-site work and remote-desktop tasks still remain with people.
Shield says GPT-6 Astra performs these tasks 35 to 40% faster on average than the earlier models it used. It also reports fewer failures and timeouts. These are vendor claims presented in an OpenAI customer story, not the results of an independent audit.
A missing API no longer has to block the workflow
For IT service providers, the appeal is mundane but real: many ticketing, device and licensing systems still force someone to click through a website. An agent that can traverse those interfaces may reduce waiting on routine requests without requiring a custom integration for every portal first.
Logging in is only the first step. A team still has to define permitted actions, customer-specific access, approval conditions and the outcome that actually closes a request. That operating discipline is what turns computer use from a polished demo into a usable workflow.
Faster clicking does not prove that the ticket was solved correctly
The 35 to 40% figure compares model execution speed for computer-use tasks. It does not show what share of requests the agent completes correctly without intervention or how often an interface change breaks a run. Shield reports higher customer satisfaction and fewer failures, but the published account provides no representative success rate for this browser layer.
A completed click is not the same as a completed request. A password reset may require identity checks, an employee departure may require authorization and a security incident may require escalation. The agent must stop as reliably as it proceeds.
Production logs will show whether delegation survives portal changes
The useful signals are success rates by task, technician interventions, incorrect changes and recovery time after a third-party interface changes. Separation of customer permissions and a complete audit trail will matter just as much.
Shield is also testing remote-desktop workflows. That expands both the possible utility and the blast radius. Results broken down by task, rather than one aggregate speed number, would make it much easier to see where the agent genuinely returns time to technicians.
Lilith's verdict
The agent can now cross a dusty portal, but a technician still draws red lines around the dangerous buttons. Success will be measured by how often it respects them, not by cursor speed.