OpenAI’s workspace agents: ChatGPT finally gets real work done

OpenAI just dropped something that actually made me sit up straighter at my desk. Workspace agents in ChatGPT. Not another chatbot wrapper, not a half-baked plugin store—these are Codex-powered agents that run in the cloud, automate complex workflows, and tie together tools without you babysitting every step.

Let me be clear: I’ve been burned by “AI agents” before. Most of them are just glorified macros with a chat interface. But this feels different. OpenAI’s been quietly building Codex into something that can reason about tasks, not just execute them line by line.

What are workspace agents, really?

Think of them as persistent, cloud-hosted assistants that can chain together multiple actions across different services. You tell it what you want done—say, “pull last month’s sales data from Salesforce, clean it, generate a summary report, and post it to Slack”—and it figures out the steps. It uses Codex under the hood, which means it understands context and can adapt when things go sideways.

They’re not running on your laptop. They live in OpenAI’s cloud, so they can keep working even when you close your browser. That’s a big deal for anyone who’s ever had a long-running task killed by a machine going to sleep.

Security that doesn’t suck

Here’s the part that usually makes me roll my eyes, but OpenAI actually handled it well. Workspace agents run in isolated environments with per-session credentials. Each task gets its own temporary access token, scoped to exactly what the agent needs. No broad API keys floating around. If you’ve ever had to clean up after a leaked credential, you know why this matters.

They also support audit logs and granular permissions per workspace. You can let the marketing team’s agent access HubSpot but not the billing system. That’s basic, sure, but it’s refreshing to see it baked in from day one instead of bolted on later.

Where they shine (and where they don’t)

I’ve been running a few test workflows. The standout use case so far is data pipeline work—pulling from databases, transforming CSVs, pushing to dashboards. It handles that beautifully. I also tried a multi-step approval flow: draft a document, send for review, collect feedback, update, and notify the team. It worked, but it was slower than I’d like for something that needs real-time interaction.

Where it falls apart? Anything requiring subjective judgment or creative nuance. If you ask it to “write a friendly email to a disappointed customer,” the result is competent but soulless. That’s not a knock—it’s an agent, not a person. But know its limits.

Pricing is another sticking point. Workspace agents run on a consumption model, and complex workflows can chew through tokens fast. I burned through $12 in credits testing a single multi-step pipeline. For a small team, that adds up. OpenAI needs to offer predictable pricing if they want enterprise adoption.

The competition is real

Anthropic’s Claude has similar agent capabilities, and Google’s Vertex AI agent builder is more flexible for custom integrations. But OpenAI’s edge is the ecosystem. If your team already lives in ChatGPT and uses common SaaS tools, the setup friction is nearly zero. You don’t need a PhD in API documentation to get started.

Still, I wish they’d open-sourced the agent framework or at least provided more detailed documentation on failure modes. When an agent errors out mid-workflow, the error messages are cryptic. “Task failed: unexpected state” doesn’t help anyone debug.

Bottom line

Workspace agents are a genuine step forward, not just another feature drop. They’re practical, reasonably secure, and actually automate real work. But they’re not cheap, not perfect, and not for every team. If you’re running repetitive, multi-step processes across a few tools, give them a shot. Just keep your credit card handy.