ServicesBlogPricingContactContact Now
AI5 min read

Shield hands legacy portal work to GPT-6 Astra agents

Shield hands legacy portal work to GPT-6 Astra agents

Introduction

For small and medium-sized businesses, the outsourced IT provider is the team that handles everything from a locked-out employee to a firewall misconfiguration that took the office offline. Much of that work happens inside portals — ticketing systems, device management consoles, software licence dashboards — that were never designed to be automated. Shield, which builds an AI operating system for IT services providers called Forge, says it has found a way to push more of that work onto agents using OpenAI's GPT-6 Astra.

The company describes Forge as a platform that combines business data with customer context so providers can complete support tasks end to end. The gap it has been trying to close is access: an agent that cannot log into an old portal is an agent that stops mid-task and waits for a human. With GPT-6 Astra handling supported browser work, Shield says technicians can now hand off those steps instead of logging in and finishing them by hand.

Why legacy portals are the real bottleneck

Technicians spend a large share of their day inside portals that manage tickets, devices and software licences. Many of these systems have no API, which means any automation has to interact with the screen the way a person would: read it, click, and type. Before the current model, Shield says its Forge agents could get stuck at exactly those systems, and the customer waited until a technician completed the step manually.

Shield had previously experimented with computer-use capabilities built on earlier models and found them neither reliable nor fast enough for production work. Mayank Jain, Head of Technology at Shield, makes the point that the systems lacking APIs are usually the oldest and least polished ones — hard to use even for a human operator. That combination of fragility and business criticality is what makes the portal problem so stubborn for managed service providers.

Speed is the commercial argument. Jain notes that when a customer is locked out of a computer or a phishing attack is underway, the priority is resolving or escalating as fast as possible. Every minute an agent is blocked on a login screen is a minute added to that clock.

How the handoff actually works

In Forge, an AI Engineer component assigns work to a computer-use agent, and GPT-6 Astra navigates the interface to complete supported tasks. The important detail for security-minded readers is the control layer: technicians decide which actions an agent may take on its own and which require human review or approval before anything is committed.

Shield says these browser workflows are live, running thousands of times each month across its IT providers. Jain describes the change as agents being able to embody what a technician does — logging into portals and legacy websites without API access and performing tasks that, in his words, only a human could have done six months earlier.

Not everything is delegated. Phone calls, on-site work and remote-desktop tasks still require human involvement, and Shield says it is currently testing remote-desktop workflows that would extend agents to more complex actions directly on a machine's desktop.

Measured gains and the limits that remain

Shield reports that GPT-6 Astra was 35-40% faster on average than the legacy models it had used for computer-use tasks. It also claims faster execution overall, fewer failures and fewer timeouts, plus less time spent by its teams writing detailed instructions for an agent or walking it through a task. On the customer side, Shield says satisfaction is higher on tickets where AI is involved, with speed of resolution cited as a critical factor.

Those figures come from the vendor and should be treated as such until independently verified. The qualitative claim is nonetheless consistent with how computer-use models have matured: the hard part is no longer recognising a button, but recovering gracefully when a page renders differently, a session expires, or a modal appears that the model has not seen before.

What IT and security teams should do now

If your organisation runs an MSP platform or is evaluating one that uses browser agents, the governance questions matter more than the model benchmark.

  • Inventory the browser-only systems. Identify which portals have no API, what data they expose, and which of them would be damaging if an agent took a wrong action inside them.
  • Scope credentials tightly. Agents that log into portals inherit the permissions of the account they use. Give them dedicated, least-privilege identities rather than a technician's own login, and keep those credentials out of shared stores.
  • Define approval gates explicitly. Decide in advance which actions — password resets, licence changes, device wipes — require a human click, and verify that the platform enforces those gates rather than merely suggesting them.
  • Log and replay. Screen-level automation should produce an audit trail detailed enough to reconstruct what an agent did, in which system, under whose authority.
  • Plan for UI drift. Vendors redesign portals without warning. Test agent workflows after upstream changes and keep a human escalation path that works when automation fails.

Grivyonx Expert Analysis

The notable shift here is not the model version but the governance layer around it. Computer-use agents effectively turn every legacy interface into an implicit API with no contract, no versioning and no deprecation notice. That is enormously useful for service providers drowning in portal work, but it moves risk from the integration layer to the identity and approval layer. The organisations that get value from this pattern will be the ones that treat an agent as a privileged, auditable operator with a narrow mandate — not as a faster pair of hands borrowing a technician's session. Shield's decision to let technicians set autonomy boundaries is the right instinct; the operational test is whether those boundaries hold under a redesigned UI and a stressed queue.

Conclusion

Shield's use of GPT-6 Astra is a practical demonstration of where agentic automation is heading: not replacing the technician, but absorbing the tedious, screen-bound steps that stall a ticket. The reported 35-40% speed improvement and the live volume of browser runs suggest the approach has moved past the demo stage for this vendor, while the remaining manual categories — calls, on-site visits, desktop work — show how much of IT support is still physical and conversational. Teams building similar capabilities should pair model selection with strict identity scoping and human approval design, which is the same discipline Grivyonx applies when embedding AI automation, DevSecOps and cloud security controls into delivery pipelines.

  • #AI agents
  • #IT support
  • #OpenAI
  • #computer use
  • #legacy systems
  • #MSP
  • #automation
Gourav Rajput

Gourav Rajput

Founder of Grivyonx Technologies