Lead
OpenAI shipped GPT-5.4 in early March 2026 with native computer-use capability — the model can now open applications, read what's on screen, and execute multi-step workflows without ongoing human input. On OSWorld-V, a benchmark designed around real desktop productivity tasks, GPT-5.4 scored 75%, just above the human baseline of 72.4%. That's not a demo. That's a general-purpose model that can work a computer at roughly human pace.
For anyone thinking about behavior agents and personal productivity, this shift is worth unpacking carefully.
What changed
GPT-5.4 introduces three things that weren't stable before:
- Native computer use: the model can parse screen coordinates from screenshots and issue mouse and keyboard commands directly — no API wrapper required
- 1M token context: long enough to plan, execute, and verify tasks across multi-step workflows in a single pass
- Autonomous offline execution: workflows can run while you're away — the model handles the chain without you watching
The practical example: pull sales figures, format them in a spreadsheet, insert into a presentation deck — in one pass, without transcription errors between steps. OpenAI is framing GPT-5.4 as the first step toward the "autonomous digital coworker."
Sources: GPT-5.4 launch — Fortune · GPT-5.4 overview — The Agency Journal
Why it matters for behavior agents
Computer use and behavior agents are adjacent but different product layers — and it's worth being clear about where each one operates.
GPT-5.4 computer use is task automation. It excels at: "complete this workflow across these apps." It's oriented toward discrete, verifiable outputs — format this, send that, compile the other thing.
A behavior agent is behavioral consistency. It's oriented toward: "help me do this thing repeatedly, across time, in the channels I actually live in." That requires:
- A habit/routine/task activity model that knows what's recurring vs. one-off
- Reminder windows (not just a fixed alarm) that adapt based on when you respond
- Done/snooze/skip event logging across sessions
- Multi-channel execution — Telegram, Slack, ChatGPT, not just a single interface
- Memory that distinguishes "inconsistent" from "abandoned" and adjusts accordingly
A model that can use a computer doesn't automatically do any of those things. Computer use is about capability. Behavior tracking is about commitment and follow-through over weeks and months.
What changes for Buffy users
The more interesting implication: GPT-5.4's computer use makes a behavior core more valuable, not less.
Here's why: as AI gets better at executing tasks, the gap that remains is behavioral — will you actually sustain the habit, maintain the routine, show up to the review? Those are the things that break in people, not in models. Powerful task automation doesn't solve the human follow-through problem.
If anything, this shift clarifies the job: GPT-5.4 handles task execution, Buffy handles behavioral execution. Plan your review in ChatGPT, have GPT-5.4 pull the data, run the habit check-in in Telegram.
One behavior core. Many capable surfaces.
What to do next
If you're thinking about how to position computer-use agents alongside habit systems: