Electric motors did not arrive in factories as a single grand machine. They arrived as replacements for the leather belts that had run off one central steam engine, and for years plants simply swapped the engine for a motor and kept the belts. The productivity gains came later, once someone rebuilt the floor plan around the fact that every station could now have its own power source. The technology was ready long before the buildings were.
I keep returning to that image when I look at what OpenAI shipped on July 10, 2026: ChatGPT Work, an agentic platform aimed at automating workplace tasks, arriving alongside the wider rollout of GPT-5. It is a serious piece of engineering, and it is also a motor being bolted into buildings that still have belts running across the ceiling.
What actually changed
The idea of an autonomous software agent is not new. It has been sitting in research papers and startup decks for years, mostly as a demo that worked beautifully once and never again. What changed in 2026 is not the concept but the substrate underneath it. Models reason more reliably across multiple steps. Tool integrations have gotten better. Enterprise data is more reachable from inside the loop.
Those three things are less exciting than a launch video, and they matter far more. An agent is, architecturally, a controller that decomposes a goal, selects tools, observes results, and revises. Every one of those steps is a place where a small error rate compounds. A model that is 95% reliable per step is a coin flip after fourteen steps. This is the mathematics that killed the last three waves of agent hype, and the reason improvements in per-step reasoning quality produce nonlinear improvements in whether an agent finishes its job at all.
So when OpenAI reports that enterprise now makes up more than 40% of its revenue and is on track to reach parity with consumer, that is not just a sales story. It is a signal that the failure rates have dropped below the threshold where a procurement committee will sign.
The adoption question is not a technology question
Survey data from June puts a useful number on the middle of the curve: 30.4% of organizations are actively developing agents with concrete plans to deploy them, on top of those already running agents in production. That is a large cohort sitting in the gap between working prototype and live system.
I would argue that gap is where the real story of the next two years lives, and almost none of it is about model capability. From my own reading of how these systems fail in practice, the blockers cluster in a few places:
- Authorization boundaries. An agent that can act needs credentials. Most enterprise permission models were designed for humans who ask before doing something unusual, not for a process that will try four hundred variations of an approach at three in the morning.
- Observability. When a multi-step agent produces a wrong outcome, the interesting question is which step went wrong and why. Most teams have logging built for request-response services, not for reasoning traces.
- Evaluation. There is no unit test for “did the agent handle this ambiguous customer request well.” Teams that succeed build evaluation harnesses first. Teams that struggle build the agent first and then discover they cannot tell whether changes made it better.
- Process legibility. You cannot automate a workflow nobody has written down. A surprising amount of agent deployment work is really business process archaeology.
From prompts to goals
The shift I find most underappreciated is the one from prompt-and-output to goal-and-decomposition. In the old interaction model, the human held the plan and the model filled in text. In the agentic model, the human specifies an outcome and the system owns the plan. That moves the skilled work from writing clever instructions to designing systems: defining what success means, what tools exist, what the agent may not do, and what happens when it gets stuck.
This is why I am skeptical of the framing that everyone will “use” agents in the way everyone uses a search engine. Search asks nothing of your organization. An agent asks you to make your processes explicit and your permissions precise. Some teams will do that work and get a genuine step change. Others will bolt the motor to the belt, automate a workflow that should have been deleted, and report modest gains.
What I would watch
Not model benchmarks. Watch whether the tooling around agents matures at the same pace as the agents themselves: permission systems designed for non-human actors, evaluation frameworks that treat trajectories rather than outputs as the unit of analysis, and interfaces that let a person inspect and correct a plan mid-execution rather than after the damage.
OpenAI is building agents for everything. Whether everyone uses them depends much less on what happens inside GPT-5 than on how quickly organizations learn to describe what they actually want done. That has always been the harder problem, and no model release solves it for you.
🕒 Published: