“GPT‑6 Astra allows our agents to embody what a technician does by logging into portals and legacy websites that might not have API access and performing tasks that only a human could six months ago,” as Mayank Jain, Head of Technology at Shield, stated. For outsourced IT service providers managing small and medium-sized businesses, support operations frequently stall on systems that lack programmatic interfaces. Technicians spend daily operational hours manually clicking through ticket managers, device registries, and software licensing hubs. The operational issue facing managers is whether autonomous computer-use agents can reliably cut manual labor across unintegrated tools without introducing costly operational errors.
Shield built Forge as an AI operating system for IT services providers, aiming to complete support tasks end to end. The architectural layer of computer-use agents bridges modern models and interface-only legacy portals without requiring custom APIs. In practice, Forge’s AI Engineer assigns incoming service work to an autonomous agent, letting OpenAI’s GPT‑6 Astra handle browser-based actions instead of forcing a technician to log in and finish them manually. As Shield explained, these browser workflows are now live, logging thousands of runs each month across its client providers.
Eliminating Interface Latency
Shield previously tested computer use using earlier models where an AI reads the screen, clicks, and types, but found them insufficient for commercial scale. According to Mayank Jain, those earlier iterations were simply not reliable or fast enough to handle intricate portals. Older corporate portals frequently lack standard design cues, creating visual hurdles that caused previous agents to stall and forced customers to wait for human intervention.
“Usually, a system that doesn’t have an API is an old-school system that also doesn’t look very nice and is very hard to use, even by a human,” Jain says.
Performance metrics shifted with the adoption of OpenAI's latest model. As Jain commented, GPT‑6 Astra was 35-40% faster on average than the legacy models Shield had used for computer-use tasks. The company reports faster overall execution, fewer workflow failures, and fewer timeouts across customer environments. The performance gain removes visual stuttering during user-interface navigation, allowing automated agents to move across legacy enterprise screens with the fluidity of human staff.
Human Supervision and Infrastructure Limits
Deploying autonomous agents into sensitive enterprise infrastructure raises the question of whether software can act without human error. Shield addresses operational risk by keeping technicians directly in the execution loop. Human specialists retain explicit control by deciding which specific actions the agent can execute autonomously and determining where mandatory manual review or approval is enforced. While browser workflows handle repetitive administrative screens, critical interactions such as phone calls, on-site interventions, and remote-desktop assignments still require direct technician involvement.
Looking beyond browser windows, Shield is testing remote-desktop workflows designed to let agents execute more complex operational procedures directly on an endpoint's desktop environment. As IT service firms face tight margins on manual support, deploying vision-guided agents directly over legacy software provides an immediate alternative to building custom API integrations, though the exact cost per transaction and the long-term inference economics of running continuous visual navigation across thousands of legacy endpoints remain to be quantified as the platform scales.
What’s behind it
What price does a business pay for having a neural network click through screens instead of building standard integrations? Browser automation absorbs part of the routine, yet leaves human engineers responsible for manual validation, customer calls, and on-site visits. According to Shield, the platform is already in production handling thousands of monthly sessions, but remote desktop scenarios remain strictly in testing. Autonomous IT support turns out to require human specialists to manually sign off on the agent's screen actions.