Work Agents

Work Agents are long-running AI agents that take a goal and carry out general knowledge work — producing documents, spreadsheets, slides, web apps and code across a person's files and tools — then return finished work for review. They grew directly out of AI coding agents, and during 2026 they became the main way the big AI labs package their models for professionals.

From Coding Agent to Work Agent

Coding was the first domain where agentic AI clearly worked. Code can be run, tests either pass or fail, and version control makes every change reviewable and reversible. Tools like Claude Code, Codex and Cursor built the machinery that long-running agents need: an agent harness that plans, calls tools, checks its own output and keeps going for hours; sandboxes; background execution; and MCP connections to outside systems.

Once that machinery existed, there was little reason to limit it to code. A spreadsheet model, a slide deck or a research memo is also an artifact that can be drafted, checked and revised in a loop. Generating a web app is coding. So in 2026 the labs repointed their coding agents at the rest of office work.

The 2026 Products

OpenAI: Codex and ChatGPT Work

On July 9, 2026, alongside GPT-5.6, OpenAI launched ChatGPT Work, a long-running work agent built on Codex that produces documents, spreadsheets, slides and web apps. According to The New Stack, Codex and ChatGPT Work together reached 8 million users by July 14, up from fewer than 1 million weekly active users in February. The combination is the clearest case of a coding agent turning into a general work agent within one product family.

Anthropic: Claude Code and Cowork

Anthropic moved in the same direction from two sides. In Claude Code's update for the week of June 29–July 3, 2026, subagents began running in the background by default, and background agents can now commit, push and open draft pull requests on their own; Claude Desktop also arrived on Linux in beta. For non-developers, Anthropic offers Cowork, which it describes as "where you hand Claude real work": "Give it a goal, and it works across your files and tools. You come back to polished work for your review." Anthropic's product page now says "Claude Cowork is now just Claude" — the work-agent mode is becoming the default way to use the assistant rather than a separate product.

Google: Antigravity 2.0

At Google I/O on May 19, 2026, Google released Antigravity 2.0 as a standalone desktop agent app with a CLI and SDK, added Managed Agents to the Gemini API, and made Gemini 3.5 Flash the default model. Antigravity began as an agentic coding environment; version 2.0 positions it as a general agent workspace.

Microsoft: Foundry to Teams and M365

At Build 2026, Microsoft moved hosted agents and the ability to publish agents to Teams and Microsoft 365 Copilot in Foundry to production. Foundry's new "Toolboxes" give agents a single endpoint for tools, MCP servers and enterprise data. Microsoft's approach is distribution: enterprises build agents once and put them where employees already work.

Browsers Fold Into the Agent

A work agent needs the web, and in 2026 the separate AI browser started to disappear into the agent app. OpenAI reportedly retired its ChatGPT Atlas browser on August 9, 2026, folding its features into a unified ChatGPT desktop app with a built-in browser plus a cloud browser for agent tasks. Anthropic took the opposite route to the same place: Claude in Chrome, an extension that lets Claude act in the user's own browser, became generally available on all paid plans on August 26, 2026. Computer use in the browser is now simply one more tool the agent can call.

That makes prompt injection the central security problem for work agents, because they read untrusted web pages and files while holding real permissions. Anthropic reports red-team attack success rates against Claude in Chrome of 16.7% for Opus 4.5, 0% for Opus 5 and Sonnet 5, and 0.3% for Fable 5, while cautioning that "prompt injection remains a moving target."

Consolidation

The shift has reshaped the companies that built coding agents. SpaceX agreed on June 16, 2026 to acquire Cursor in an all-stock deal valuing it at $60 billion, closing on August 14; Cursor said the deal gives it "access to the largest fleet of GPUs in the world." Cognition, maker of Devin, raised $1 billion at a $26 billion post-money valuation in May 2026, with a $492 million annualized revenue run-rate. Agent builders are also becoming distribution partners: Anthropic's Claude Marketplace, launched September 23, 2026, lists partner agents from Cursor, Harvey, Lovable, Snowflake and others.

Pricing Moves to Usage

Agents that run for hours consume far more compute than chat, and flat per-seat pricing has not survived that. On June 1, 2026, GitHub Copilot moved to usage-based "AI Credits" while keeping list prices unchanged; GitHub said agentic sessions had made premium-request pricing "unsustainable." Claude Marketplace lets enterprises pay for partner agents with committed Anthropic spend. The emerging model is a subscription that buys a pool of usage, with agent work drawn down against it — closer to cloud billing than to SaaS seats, and tied directly to the inference economy.

Why It Matters

Work agents change the unit of AI use from a prompt to a delegated task. The questions that matter shift accordingly: how long an agent can work reliably (the time horizons that METR measures), how it is permissioned and audited, and how it is billed. They also run on shared plumbing — MCP for tools, A2A for agent-to-agent handoffs, AGENTS.md for instructions — now governed by the Agentic AI Foundation. For the agentic economy, work agents are where agents first earn revenue at scale.