Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableOpenAI Blog→ original

Shield Hands GPT-6 Astra Work in Legacy Portals Without APIs: Agents 35–40% Faster

IT outsourcer Shield keeps part of its routine work in legacy portals with no API, meaning no entry point for programs, only forms and buttons. Previously, the Forge platform's agent got stuck there, and the client waited for a technician. Now GPT-6 Astra handles these tasks: it looks at the screen and clicks like a human, on average 35–40% faster than earlier models (estimate from Shield's chief technology officer). The absence of an API, then, is not a reason to call a person, but agent permissions should be defined in advance.

AI-processed from OpenAI Blog; edited by Hamidun News
Shield Hands GPT-6 Astra Work in Legacy Portals Without APIs: Agents 35–40% Faster
Source: OpenAI Blog. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceOpenAI Blog ↗
Editor-in-Chief
Key metric #1Thousands of runs per month at IT providers on the Forge platform
Key metric #2Technicians decide what the agent does without asking and what requires review first
Key metric #3Calls, on-site visits, and remote desktop work still remain with people

Picture a typical IT support ticket: an employee's laptop is locked, and to unlock it, someone has to log into an old web portal that runs on forms and buttons. That portal has no API, a "door for programs": a person can get in, but an automated agent cannot. Shield set out to solve this problem. The company says that after switching to OpenAI's GPT-6 Astra, its agents complete such browser tasks on average 35–40% faster than on the earlier models Shield used to control the screen. This is the company's own estimate, not an independent measurement.

Where the agent hit a wall

Shield builds Forge, an "AI-based operating system" for IT service providers. These firms serve small and mid-sized businesses: they unlock an employee's laptop, fix a misconfigured firewall that left an office offline, and so on. Forge collects data about the client and the context of the request so the agents can carry the ticket through to completion.

But technicians also have to work in old portals for tickets, devices, and licenses. According to Mayank Jain, Shield's chief technology officer, systems without an API usually have outdated interfaces, and they are awkward even for humans. Forge's agents used to get stuck in these places, and the client waited while a technician logged in manually and finished the step. When a computer is locked or a phishing attack is underway, every minute of delay is costly. Jain puts it this way: a human is needed when judgment is required, not because the agent wasn't let into the system.

Shield had already tried computer use, a mode in which the model reads the screen, moves the cursor, and types text. But according to Jain's assessment, the earlier models were not reliable or fast enough: you could see them stumble while trying to figure out the interface.

How it works now

Here is the setup in Forge. An internal "AI Engineer" receives a task and passes it to an agent with computer use. GPT-6 Astra navigates the portal's interface and performs the supported actions: it logs into an account, finds the right ticket, and changes a status or license. Technicians set the boundaries themselves: which actions the agent takes without asking, and which it must surface for review or confirmation.

This detail matters more than it seems. An agent with click access to a device and license management console can do real damage. That's why the split between "do it yourself" and "ask a human" is part of the product, not a formality.

According to Shield, these scenarios already run in production: thousands of runs per month at IT providers that use the platform. Jain says GPT-6 Astra is on average 35–40% faster than the models the company previously used for computer use. He also notes fewer failures and timeouts, and that teams spend less time writing detailed instructions for the agent and guiding it step by step. In his account, customer satisfaction is higher on tickets where AI is involved, and the decisive factor is resolution speed. Jain's most vivid impression: before, you could see the model struggling with the interface, and now it looks like watching a person click through a screen.

What to keep in mind

This is a marketing case study on OpenAI's website, and all the figures are the words of a client representative. There is no measurement methodology: it's unknown how the "35–40%" was calculated, on which tasks, or what the "earlier models" were. There is also no cost data: how many tokens and screenshots a task consumes, and what a single pass through a portal costs. For evaluating the economics, these are the most important numbers, and they will have to be measured on your own workflows.

There are also honest limits. Phone calls, on-site visits, and tasks through a remote desktop still require a human. Shield is testing remote desktop scenarios, where the agent could perform more complex actions directly on the desktop, but this is still an experiment.

Practical takeaway

The lesson for developers and product managers is that having an API is no longer a prerequisite for automation. If your process includes an "old portal that the intern usually logs into," it is a candidate for a computer-use agent. Start with narrow, repeatable tasks, define action permissions in advance, keep human confirmation for irreversible steps, and collect your own metrics: execution time, failure rate, timeouts, and cost per run. Only then can Shield's claim be tested against your numbers rather than the numbers in a press release.

Story

Key Questions Answered

Where the agent hit a wall?

Shield builds Forge, an "AI-based operating system" for IT service providers. These firms serve small and mid-sized businesses: they unlock an employee's laptop, fix a misconfigured firewall that left an office offline, and so on. Forge collects data about the client and the context of the request so the agents can carry the ticket through to completion. But technicians also have to work in old portals for tickets, devices, and licenses.

How it works now?

Here is the setup in Forge. An internal "AI Engineer" receives a task and passes it to an agent with computer use. GPT-6 Astra navigates the portal's interface and performs the supported actions: it logs into an account, finds the right ticket, and changes a status or license. Technicians set the boundaries themselves: which actions the agent takes without asking, and which it must surface for review or confirmation. This detail matters more than it seems.

What to keep in mind?

This is a marketing case study on OpenAI's website, and all the figures are the words of a client representative. There is no measurement methodology: it's unknown how the "35–40%" was calculated, on which tasks, or what the "earlier models" were. There is also no cost data: how many tokens and screenshots a task consumes, and what a single pass through a portal costs. For evaluating the economics, these are the most important numbers, and they will have to be measured on your own workflows.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Private AI Builders Community · Vibe Club

Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews

Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.

Join Vibe Club →Explore Hamidun Academy
What do you think?
Loading comments…