GPT-5.4 Gets Computer Use

Edition 164 | March 9, 2026

You know you’ve built a great agent when it has trust issues with itself.

Welcome back to Building AI Agents, your biweekly guide to everything new in the field of agentic AI!

In today’s issue…

• OpenAI releases GPT-5.4 with native computer use

• Anthropic's labor data reveals AI impacted industries

• Google makes Gmail, Drive, and Docs agent-ready

• OpenAI launches Codex Security in research preview

• LangChain CEO: better models won't get your agent to production

…and more

🔍 SPOTLIGHT

When Anthropic released its new labor market research report last week, the discourse followed the usual script. Headlines focused on whether AI is coming for your job, pundits weighed in with familiar predictions about mass displacement, and the takes machine cranked out another cycle of doom-and-reassurance. But the most important finding in the report isn't about job loss at all. It's that the biggest opportunity in AI agents right now may sit in the gap between what models can already do and what businesses have actually operationalized.

The report introduces a new metric the researchers call observed exposure: a measure of how much of a given job's tasks are not just theoretically possible for an LLM to handle, but are already showing up as real, automated, work-related usage on Anthropic's own platform. The distinction matters. Researchers have already mapped where language models should be useful on paper, especially in fields like software, finance, and office administration. Anthropic's contribution is showing how little of that potential has actually made it into production. In Computer & Math occupations, for instance, theoretical exposure sits at 94%. Observed coverage? Just 33%. The pattern repeats across nearly every category. AI could be doing far more than it currently does, and the bottleneck is not capability, it's deployment.

That alone would make the report worth reading. But there is a second finding, quieter and arguably more telling. The researchers found no systematic increase in unemployment for workers in the most AI-exposed occupations. No mass layoffs, no white-collar apocalypse. What they did find was early, tentative evidence that something subtler is happening. Among workers aged 22 to 25, the rate of entry into high-exposure jobs appears to have dropped roughly 14% compared to 2022. The result is barely statistically significant, and Anthropic is careful to say so. But the pattern is consistent with what several other research groups have found independently. Most notably, a Stanford Digital Economy Lab study using ADP payroll data from millions of U.S. workers attributed the decline primarily to slowed hiring, not increased firings. The first labor market effect of AI may not be people losing jobs. It may be jobs quietly never getting filled.

This is what early-stage automation actually looks like in the real economy. Not departments being gutted overnight, but a company deciding not to backfill the junior analyst role. A support team running leaner because a chunk of ticket volume now routes through software. A finance team that used to onboard two new hires a year now onboarding one, with an automated workflow picking up the slack. The report's most exposed occupations (computer programmers, customer service representatives, data entry keyers, financial analysts) all share a common profile: the work is screen-based, language-heavy, modular, and routed through defined software environments. In other words, the exact shape of work that agents can absorb piece by piece.

And this is where I believe the report becomes less of a policy paper and more of a builder's roadmap. The gap between theoretical capability and actual deployment is not a warning, it is a to-do list. Every point of distance between the blue and red bars on Anthropic's charts represents a workflow that an AI could handle, that someone out there probably needs handled, and that no one has yet turned into a dependable system. The occupations at the top of the exposure list are not there because they are doomed. They are there because the work is already structured for automation, and still far from fully operationalized.

Anthropic did not write this report for agent builders or company owners. But if you are one, it might be the most useful thing published this year.

As always, keep learning and building!

—AP