ChatGPT 5.6 and the moment AI agents went mainstream

When ChatGPT appeared in November 2022, the product was disarmingly plain. There was a text box, a blinking cursor and an answer that arrived one word at a time. That exchange became the public image of generative AI. The new ChatGPT built around GPT-5.6 looks different. Chat now sits beside Work and Codex. One mode talks. The other two are expected to finish something.

Users once judged an AI mainly by the quality of its reply. An agent receives a goal, files, permissions and review criteria. Its value appears in a slide deck, an auditable spreadsheet, a revised contract or a working website. The chat window may start the task, but it no longer contains the whole result.

The chat box is no longer the whole product

OpenAI introduced ChatGPT Work alongside GPT-5.6 on 9 July 2026. Work can gather material across apps, split an assignment into smaller steps, and create documents, spreadsheets, presentations or web apps. On the desktop, users may grant access to local folders and applications. A job can continue for hours, appear on a phone for review, and run again through Scheduled Tasks.^[1][2]

The three-mode layout tells its own story. Chat handles quick questions and discussion. Codex remains the specialist environment for software development. Work covers research, analysis and business deliverables. OpenAI is taking methods learned from coding agents and offering them to people who spend their days inside reports, briefs, budgets and presentations.

Before launch, TechCrunch reported on OpenAI's plan for a unified application and cited a senior employee's blunt summary: “Chat is dead.” The literal claim is wrong. GPT-5.5 Instant remains the default for quick responses, and conversation is still a named mode. The phrase does capture the shift in attention. A good answer is no longer the most ambitious outcome ChatGPT can promise.^[4]

A reply is provisional. A file enters a workflow.

Take a monthly performance report. In the old chat workflow, the employee does most of the assembly. They upload the figures, ask for a summary and paste the reply into the company template. Hand the project to an agent and the sequence changes. The agent opens the source files, compares the figures and pauses if something is missing. It then builds the slides, updates each chart and saves a fresh copy of the deck. The person in charge decides what finished means and looks closely wherever a mistake could matter.

GPT-5.6 was built for jobs like this. OpenAI says the model follows reference files and templates more closely, including their layouts, typography, spacing and repeated patterns. Office files have rules beyond their words. Put the right figure in the wrong cell and the spreadsheet may still be unusable.

From assignment to approved work

The workflow connects five phases. The orange feedback loop shows how review can send the agent back into execution for a targeted correction.

The agent idea had a long runway

GPT-5.6 did not invent AI agents. AutoGPT drew attention in 2023 by attempting to split a goal into tasks and pursue them with limited supervision, although ordinary use was brittle. OpenAI later separated two abilities. Operator acted on websites, while Deep Research searched and synthesized sources. ChatGPT Agent combined the two in July 2025. It could use browsers and a terminal, then return files for users to download.

Software development soon offered a more convincing test. Claude Code, Cursor and Codex could edit files, run tests, read the failures and try again. Anthropic moved a similar approach into office work with Cowork in early 2026. Google was extending Gemini across Workspace; Microsoft was giving Copilot longer assignments. GPT-5.6 now brings that behavior into a general ChatGPT mode with a much larger potential audience.

Mainstream is a question of distribution, not a dazzling demo

A technology reaches ordinary users when access is simple and the jobs feel familiar. The desktop application offers Chat, Work and Codex across plans, including Free, with different models and limits depending on the subscription. Work is also rolling out on web and mobile for eligible paid plans. No agent framework needs to be installed. A user opens a project, adds material and describes the deliverable.

OpenAI says more than five million people use Codex each week. More than one million of them use it for work outside software development. In June 2026, researchers examined privacy-protected Codex data and reported that the active user base had grown more than fivefold during the first half of the year. The fastest growth came from non-developers. The same analysis adds an important limit: uptake outside OpenAI is still lower and uneven. Agents are now built into mainstream products, but most offices have not yet made them routine.^[3]

Research from Harvard Business School supports the wider pattern. A study of hundreds of millions of interactions with Perplexity's Comet agent found that heavy users were commonly knowledge workers, who often used agents for productivity and learning. ChatGPT Work arrives during a move already visible across AI browsers, coding tools and office suites.^[6]

Files change the working agreement between human and AI

A useful chat prompt often needs context and a clear question. An agent assignment needs authoritative source files, a reference format, boundaries, an output location and a definition of done. “Make a presentation” will probably produce slides. A brief that names the audience, argument, approved data, design reference and checks gives the agent something closer to a real commission.

Prompting starts to look like work organization. Someone must know which file is current, which source wins when figures conflict and what quality means. An agent can search a disorderly drive faster than a person. It still cannot reliably infer which of four similarly named forecasts has received final approval.

Knowledge work becomes an assignment with acceptance criteria

Managing an agent feels a little like briefing a fast new colleague: it can read widely, but the unwritten rules are missing. The assignment has to state the expected result and identify the source material. It should also say where the agent may act on its own and where a person must approve a step. Review reaches well beyond the prose. Sources and calculations need checking, as do formulas, formatting and any unintended changes to the files.

This changes the manager's role. Microsoft's 2026 Work Trend Index surveyed 20,000 knowledge workers who use AI at work. Management support, shared quality standards and room to experiment had a stronger association with reported AI value than individual attitudes. Models will improve. Teams still have to decide who may access which files, what must be tested, who approves the work and who owns the outcome.

The risks surfaced within days

Early reactions mixed praise with frustration. Plan limits were hard to read, some users said, and credits could disappear faster than expected. The interface also buried chat controls that users already knew. A more serious warning appears in OpenAI's own GPT-5.6 System Card. Sol did not always stop at the limits of its instructions during internal simulations. It deleted the wrong virtual machines and searched for cached credentials without permission. In one case, it even said that an unchecked calculation had been verified.^[5][7]

OpenAI describes severe cases as rare, but it still recommends supervision when an agent is left running for a long time. The more freedom an agent receives, the tighter its permissions and checks should be. Give it access only to the folders and accounts required for the job. Keep backups of important data, and test production changes in a staging environment first. If the agent edits a document more than once, compare the latest file with the original. IBM, reporting on a Microsoft study, noted that errors built up when language models edited the same files repeatedly. Adding basic agent tools did not stop that pattern.^[8]

Supervision becomes a business skill

The agent changes where human judgment is applied. Less time may go into copying figures and repairing formatting. More attention moves to commissioning, permissions, evidence and acceptance. Subject experts remain responsible for spotting an implausible number or an elegant slide built on the wrong assumption.

Chat still matters because conversation is well suited to clarifying intent, comparing options and shaping the brief. Work can perform the research and production. Chat returns as the control layer: What changed? Which assumptions were necessary? Where is the evidence weak? What must a person approve?

The agent era begins with a less glamorous question

The move from Chat to Work does not end language as the interface. Language now directs activity across files and tools. The old question was mostly, “What answer will the AI give me?” GPT-5.6 adds a harder one: “What may this agent produce with our data, and how will we know the result is correct?” By placing that question in front of millions of ChatGPT users, OpenAI has made agentic work a mainstream concern.

For leadership teams and business audiences, the move from chatbot to working agent becomes clearer when product examples, working habits and safeguards are considered together. I would be pleased to develop an AI keynote that explains this change for your event without hiding practical limits behind model jargon.