GPT-6 Astra Computer Use: Job Hunt Guide

GPT-6 Astra computer use, described in a September 2026 video, can research jobs, apply, follow up, and network on LinkedIn while you track every role in a spreadsheet. The workflow behind it has four steps: build context in a project, write an SOP file, package it as a reusable skill, and swap screen-driving steps for CLI tools to cut token cost.

ChatGPT-6 Astra computer use is and what it did in this workflow

GPT-6 Astra computer use, as described in a September 11, 2026 video by the channel Eric Tech, is an agent mode that visually controls your computer to run a job search end to end. The video's author, who describes himself as a senior AI engineer and former Amazon and Microsoft employee, says he used it to research roles, apply, send follow-ups, and message people on LinkedIn while tracking everything in Google Sheets. Every capability described below is speaker-reported from that video, not independently verified by this article.

The demonstrated loop had four parts: build context about the job seeker in a project, interview the user into a written standard operating procedure, package the workflow as a reusable skill, and replace expensive screen-driving steps with CLI tools and API calls. The same pattern applies to any repetitive desktop task, not only job hunting.

A useful caveat up front: computer use agents that fill forms, send messages, and submit applications on your behalf can violate the terms of service of platforms like LinkedIn, and a wrong click can submit a bad application. Review each run before trusting the output at scale.

How computer use works, and how it differs from MCP

Computer use lets an AI agent operate a computer visually instead of calling applications through code. In the workflow described by the video, the agent takes a screenshot of each screen, sends it to the model, and the model decides the next mouse click, keystroke, or scroll. That means it can drive desktop apps, browser apps, and anything a human user could reach.

The video contrasts this with MCP, the Model Context Protocol, an open standard that lets agents control applications programmatically through defined tools. Computer use is broader but slower and more error-prone; MCP and direct API calls are narrower but cheaper and more stable, which matters in the CLI step later.

The speaker also claims OpenAI shipped an earlier computer use capability on April 23 of this year alongside what he calls the GPT-5.5 model, and that the newer model is about 70% more token efficient. Both figures come from the video itself; this article could not verify them against OpenAI's primary sources, so treat them as speaker-reported.

Step 1: Build the context in a Codex project

Before the agent can fill applications or send messages, it needs to know who you are and what you want. The video's author does this inside OpenAI Codex, the coding agent tied to a ChatGPT account. In the Codex interface he opens the projects tab, creates a new project, and points it at a desktop folder, keeping one project per automation: one for job finding, one for LinkedIn networking, one for event outreach.

In that project he starts a session and runs an initialization command, then uses what he calls a grooming skill: the agent relentlessly interviews him about the goal, the workflow, decision trees, and edge cases until both sides reach a shared understanding. The output of that interview is a Markdown file containing the full SOP, which becomes the single source of truth for every later run.

Step 2: Test the SOP with computer use

With the SOP written, the author runs it in a fresh session and lets computer use execute it. In the video, the agent reads his Google Sheets tracker, searches job boards in the browser, and applies to five roles. When he changes direction and asks for jobs at recently funded YC companies, the agent re-searches and applies again under the new constraint.

This test run is where edge cases surface: login walls, unusual application forms, and ambiguous instructions. Fixing the SOP file at this stage is cheap; fixing it after the workflow is packaged and reused daily is not.

Step 3: Package the workflow into a one-command skill

A skill turns the tested SOP into a reusable command so you never re-type instructions. The author asks the agent to convert everything discussed in the session into a skill named job search, then triggers it with a slash command in a fresh context window. On that run he reports the agent added 15 companies to the tracker, applied to those jobs, and generated a tailored resume for each one.

The value is consistency: each run starts from the written SOP rather than from a remembered conversation, and the same trigger can be reused for networking outreach or follow-up scheduling. Any agent platform with a skills or custom-command feature supports this pattern, including Claude Code, Anthropic terminal-based coding agent.

Step 4: Swap screen-driving steps for CLI tools and API calls

Even with the reported 70% token-efficiency gain, pixel-driven computer use costs more tokens than structured calls, because every screenshot consumes input. The author's final step is to ask the agent to audit its own skill and move steps that do not need a screen into CLI commands or API calls, reserving computer use for pages and forms that genuinely require visual interaction.

In the video this produced a small jobs CLI with subcommands for context, inspection, and tracker plans. The agent reported measured savings of roughly 22% for the job-detail reading step, where the workflow fetches listing content and decides whether to apply. That 22% is speaker-reported and specific to that step and setup, not a general guarantee for every workload.

The trade-off is stability versus coverage. User interfaces change with every product update, so visual automation breaks more often; scripts and APIs change less frequently but only cover steps that expose an interface. The video's hybrid approach uses each for what it does best.

Comparison: computer use vs. MCP vs. CLI tools

The three execution modes in this workflow differ in cost, Stable Diffusion coverage. The table summarizes how the video's author uses each one.

Mode	What it drives	Token cost	Best use in this workflow
Computer use	Any visible screen, via screenshots	Highest	Login walls, web application forms
MCP	Apps exposing protocol tools	Medium	Structured app control where a connector exists
CLI / API calls	Scripts and endpoints	Lowest	Search, listing reads, tracker updates

The practical rule from the video: default to the cheapest mode that can complete a step, and fall back to computer use only when a screen is unavoidable.

FAQ

• Do I need GPT-6 Astra specifically to automate a job hunt this way? The four-step pattern works with any computer-use agent plus a coding agent that supports projects and custom skills. The video uses OpenAI's tools, but the structure, context file, SOP, skill, CLI conversion, transfers to other agents.

• Is the 22% token saving a guaranteed figure? No. It is the video author's own measured saving for the job-detail reading step in his workflow. Your savings depend on which steps you move to CLI tools and how the agent is configured.

• Can an agent apply to jobs on LinkedIn without violating rules? The video shows LinkedIn messaging and connection requests running automatically. Automated activity can conflict with a platform's terms of service, so check the current terms and rate-limit the agent before running it unsupervised.

• Why write an SOP file instead of just prompting each time? The SOP file makes the workflow reusable in fresh context windows, keeps decision rules consistent between runs, and gives you a single document to edit when a step breaks.

• What should computer use be reserved for once a CLI exists? Screens and forms with no clean interface: login flows, employer-specific application forms, and any page where a structured call is unavailable.

Source video