ChatGPT agent mode explained: What it can and cannot do

TL;DR:

• ChatGPT Work, a unified mode OpenAI launched on July 9, 2026, handles multi-step tasks and delivers finished files instead of chat replies.

• The gaps are scale and reliability: native connectors may require per-action confirmation for write operations and session-level re-authentication, and the depth of integration is uneven across that directory.

• Composio is action infrastructure for knowledge work agents: one SDK to act across 1,000+ apps & integrations, learning from every run, so your app connections stay live, your workflows execute reliably, and your agent gets better over time.

Copying campaign data from ChatGPT into your CRM by hand burns hours every week on a problem that should already be solved. ChatGPT's agentic capabilities have matured, but the gap between what the native product can do and what your marketing stack needs is still large enough to matter.

We'll walk you through what ChatGPT Work (which replaced the earlier Agent Mode feature) can genuinely handle, where it stops, and how we help you fill those gaps without building anything from scratch.

What is ChatGPT Work and how does it function

ChatGPT Work does not answer a question and stop. It takes a brief, works in the background for minutes or hours, and hands you a finished file: a spreadsheet, a slide deck, a report, or a working web app.

Agent mode vs. standard chat workflows

Standard chat typically works as a single-turn exchange. You send a message, ChatGPT replies, and the interaction ends. ChatGPT Work runs what appears to be a continuous loop: Plan, Execute, Observe, Refine. It writes a plan, takes an action, reads the result, adjusts, and keeps going until the goal is met or it hits a blocker.

The practical difference matters: asking ChatGPT to "write a product launch email" is a single-turn task. Asking it to "research our three top competitors, summarize their positioning, pull our existing launch brief, and draft five email variations with subject line tests" is a Work-mode task. One critical constraint: actions that could affect your accounts or spend money (like signing into services or making purchases) reportedly pause and hand control back to you. You are supervising, not fully delegating.

Triggers for ChatGPT agent mode

ChatGPT Work activates when a task requires more than one distinct step and produces a finished deliverable rather than a conversational reply. Concrete triggers include:

• Complex research pulling from multiple web sources

• Data analysis and transformation on uploaded files

• Multi-step document creation (reports, briefs, slide decks)

• Tasks requiring multiple web interactions

• Code generation and execution in its Python sandbox

Single-question lookups, quick rewrites, and brainstorming sessions stay in Chat mode. Inside ChatGPT's interface, a mode selector at the top lets you switch between Chat and Work.

Key capabilities of ChatGPT agent mode

ChatGPT Work ships with a meaningful set of built-in tools. Within its environment, the native capabilities cover most research, file, and scripting tasks a marketing team runs weekly.

Creating workflows for ChatGPT agents

Setting up a multi-step task in ChatGPT Work typically follows three steps:

1. Select Work mode: Open the mode selector at the top of the ChatGPT interface and switch from Chat to Work.

2. Define your goal: Write a specific task description including what you want produced, what inputs exist (uploaded files, URLs), and any format constraints.

3. Review and confirm: ChatGPT Work presents its plan before executing major actions. Read it, confirm, and check back when it surfaces a decision point.

The clearer the goal, the fewer interruptions you get. A task like "analyze the attached Q2 campaign report and write a one-page executive summary with three recommendations" runs cleanly. Vague prompts produce vague plans that stall immediately.

Managing data within ChatGPT agent mode

ChatGPT Work handles uploaded files well. It reads documents, extracts structured data, reorganizes spreadsheet columns, works with table data, and outputs transformed versions of what you give it. The table below shows what changes for a common task:

ChatGPT Work can clean your data inside its environment. Native write-back to HubSpot is possible through the built-in connector with user approval settings, and that connector is available to all HubSpot customers across all tiers with a paid OpenAI subscription (Plus, Team, Pro, Enterprise, or Edu). For multi-app workflows running reliably in the background across 1,000+ apps, that's where we come in.

What ChatGPT Work cannot do natively

The capabilities above cover what ChatGPT Work does well in isolation. The problems start when your workflow requires the agent to interact reliably with the apps you use every day.

Linking your current workflow apps

ChatGPT Work does have OAuth-based connectors (connections that let ChatGPT access your accounts after you sign in) for apps like Gmail, Google Drive, and Salesforce. These connectors work for basic, session-level access. Two limitations matter for production use.

First, session-level auth may mean the connection doesn't persist reliably between sessions without the user re-authenticating. Second, while ChatGPT Work can @-mention apps from a directory of 1,400+ connectable apps, the depth of integration and write capabilities are uneven across that directory.

People who need broader coverage try building Custom GPT Actions with a direct API connection (a way to connect ChatGPT directly to an app's backend). This works for a single app in a controlled environment but requires manual credential management and doesn't scale across multiple apps. When access permissions expire, the connection may fail without clear notification.

Check your ChatGPT Work settings menu for native integrations and use them where they cover your need. For anything requiring persistent, multi-app access, we built Composio to handle the connection layer so ChatGPT can focus on the reasoning task.

Limitations on long-term memory

As of its July 9, 2026 launch, OpenAI describes ChatGPT Work as an agent for longer, more involved tasks that can work across connected apps and files and remain on complex projects for hours.

As a session grows, early instructions drift out of the active memory as new content is added. A campaign brief you referenced at the start of a long session may no longer be in scope by the time the agent writes the fourth content variation. As stated in OpenAI's July 9 launch announcement, ChatGPT Work began rolling out on web and mobile to Pro, Enterprise, and Edu plans, with Plus and Business access following over the next few days. On the ChatGPT desktop app, Chat, Work, and Codex were made available globally across all plans, including Free, on Windows and Mac. Where available, tasks can execute without you manually opening a new session. However, between scheduled executions, the system is not actively monitoring your inbox or analytics dashboard in real time.

Why ChatGPT cannot pull live metrics

Live API data from Google Ads, HubSpot, or any login-protected source requires connection management that ChatGPT Work doesn't handle natively between sessions. Access permissions from providers like Google expire after 60 minutes. When ChatGPT Work attempts to pull data requiring a Google Ads connection between sessions, the expired permission may cause the request to fail or prompt you to sign in again mid-task. Neither works for a reporting workflow that should run in the background.

We solve this with a managed token refresh layer (a system that automatically renews access permissions before they expire). For marketing teams that need live campaign data in their agent workflows, this is the difference between a workflow that runs once and one that runs every morning.

Optimizing agent mode for maximum daily output

Given the native constraints above, here's how to get the most from what ChatGPT Work does well while avoiding the failure modes that waste time.

How to scope requests for better output

ChatGPT Work performs best on bounded tasks with clear inputs and defined outputs. Break a six-week campaign strategy into components:

• Session 1: Competitive research on three named competitors, output a comparison table

• Session 2: Audience brief based on uploaded persona document and research output

• Session 3: Content calendar draft for one channel using the audience brief

Each session produces a finished artifact you can review and hand to the next session as input.

Confirm quality before sharing results

ChatGPT Work will deliver a finished file with confidence regardless of whether the underlying data was correct. Before sharing any output, cross-reference one set of data points against your source of truth, check that cited figures match what you provided, and run any generated email copy through your own brand voice filter. This takes five minutes and prevents the credibility issues that come from shipping stale information to a client or executive.

How to frame prompts for agent mode

Moving from open-ended prompts to task-definition prompts is the single highest-leverage change you can make. These templates work consistently:

For research tasks: "Research [topic]. Focus on [specific angle]. Use [URL 1], [URL 2] as primary sources. Output a [format] with [X] sections covering [A], [B], [C]. Maximum [word count]."

For file analysis: "Analyze the attached [file name]. Identify [specific patterns or issues]. Output a table with columns [X], [Y], [Z]. Flag any row where [condition]."

When to use agent mode vs. standard ChatGPT

Most tasks do not need agent mode. Knowing when to switch saves you from burning through your monthly agent message quota on tasks that standard Chat handles in ten seconds.

Ideal workflows for ChatGPT agent mode

Work mode fits when the task has multiple steps, requires gathering or transforming data, and ends in a deliverable you can hand off or publish. The three highest-value use cases for marketing teams are:

1. Content repurposing: Upload a webinar transcript or long-form article, prompt Work to extract key insights, and rewrite them as five LinkedIn posts, a summary email, and a FAQ document.

2. Campaign reporting: Upload raw data exports from your ad platform and CRM, prompt Work to build a summary with performance tables and trend commentary.

3. CRM data hygiene: Upload a contact export, prompt Work to identify duplicates, flag missing fields, and standardize formats. Combine with our HubSpot toolkit to write the cleaned records back directly at scale.

Choosing speed over agent autonomy

Standard Chat is faster for tasks with a single, clear output: writing one email, brainstorming headline options, summarizing a document you paste in, or generating interview questions. Work mode adds planning overhead that isn't worth it for a five-minute task. The practical rule: if the task would take you under 15 minutes manually, use Chat. If it would take you 30 minutes or more, hand it to Work mode and let it run while you do something else.

Connect ChatGPT Work to your apps

The missing execution layer

ChatGPT can reason about what needs to happen, but without a managed action layer it can't reliably execute across your marketing stack. Composio is action infrastructure for knowledge work agents: one SDK to act across 1,000+ apps, learning from every run. Across more than 1M accounts, it provides one governed path to act on Gmail, HubSpot, Google Ads, and your entire workflow, without you building or maintaining each integration.

• Integrations: One governed path to act across 50,000+ agent-ready tools, with structured schemas formatted for LLM consumption. ChatGPT gets clean, actionable data from every tool call without you mapping APIs or maintaining connectors.

• Auth: Your workflows run without manual sign-ins or re-authentication interruptions. Composio provides one governed auth path across every connected system, managing OAuth 2.0, API keys, and JWT tokens with automatic refresh, so access permissions never expire mid-run. Composio is SOC 2 and ISO 27001-certified.

• Execution: Tool calls route through the execution layer with error handling and retry logic built in. When an API call fails, it handles the recovery so your workflow doesn't break.

• Self-Learning: The system processes 300M+ tool calls per month and uses that signal to continuously refine the schemas agents see, tracking which tool descriptions and parameter formats produce successful calls and adjusting automatically. The result is a 30% accuracy improvement and 2x fewer tokens consumed per tool call.

This creates a useful division of responsibilities. ChatGPT Work provides the reasoning, planning, context gathering, and artifact creation, while Composio provides the integrations and actions needed to interact with external systems. Together, they can turn Work from an assistant that primarily works with the information available to it into an agent capable of coordinating workflows across a much broader software stack.

Get started: Composio's free tier includes 20,000 tool calls per month with no credit card required, which covers meaningful evaluation for a marketing team running a few connected workflows. The $29 per month plan adds 50,000 tool calls.

FAQs

Can ChatGPT Work connect to HubSpot or Salesforce natively?

Yes, through built-in connectors, but with limits. For multi-app workflows running reliably across 1,000+ apps with automatic execution, you need a managed auth layer like Composio.

How long does it take to connect an app like Gmail to a ChatGPT agent using Composio?

Based on user-reported timelines, Gmail and Google Drive connections through Composio take under 30 minutes from setup to first authenticated tool call. The free tier covers initial evaluation with 20,000 tool calls per month and no credit card required.

Key terms glossary

MCP (Model Context Protocol): Think of it as a universal connector for AI. MCP is a way for AI assistants to connect to external tools and data sources in a standardized way.

OAuth: A security standard that lets apps access your account without seeing your password. When you "Sign in with Google," that's OAuth.

Token refresh: The automatic process of renewing access permissions before they expire, so your integrations keep working without you signing in again.

API (Application Programming Interface): A way for two applications to talk to each other. When ChatGPT pulls data from HubSpot, it uses HubSpot's API.

Context window: The amount of text an AI model can "see" and work with in a single conversation. Measured in tokens (roughly 0.75 words per token).