AI agents went from research demos to everyday software between 2024 and 2026. Assistants from Anthropic and OpenAI can operate a computer, automation tools such as Zapier sell agents that act across thousands of apps, and the largest AI companies now back shared standards for connecting agents to other software. If you want to know what comes next, three trends you can already measure are a better guide than any single product launch: how long a task an agent can finish, how easily it can reach your tools, and how much trust companies and regulators give it.
Agents can finish longer tasks each year
The AI evaluation group METR tracks the length of task, timed by how long a skilled person takes to do it, that AI models can complete with 50% reliability. Its March 19, 2025 study found that this length had doubled about every seven months for six years. An updated version of the method, published on January 29, 2026, put the doubling time for models released since 2024 at about 89 days, roughly three months.
Two caveats come with those numbers. A 50% success rate means a model still fails half the tasks at the edge of its ability, so the length of task an agent can do reliably is much shorter. METR also says the trend is sensitive to which tasks it uses. Even so, its 2025 paper extrapolated that within a decade, agents could handle a large share of software tasks that now take people days or weeks.
In practice, the size of job worth handing to an agent keeps growing. A task that needed a person to break it into ten small requests in 2024 may run as one request a year or two later.
Shared standards make agents easier to connect
An agent is only useful if it can reach your apps. When Anthropic released MCP, it pointed out that every new data source needed its own custom connection, which made connected AI systems hard to scale.
The Model Context Protocol (MCP), released on November 25, 2024, is an open standard for connecting AI apps to tools and data. On December 9, 2025, the Linux Foundation launched the Agentic AI Foundation with MCP as a founding project, alongside OpenAI's AGENTS.md instructions format for coding agents and Block's goose agent framework. Its platinum members include Amazon Web Services, Anthropic, Google, Microsoft and OpenAI. At launch the foundation counted more than 10,000 published MCP servers, and Claude, ChatGPT, Microsoft Copilot, Gemini, Cursor and VS Code all supported the protocol.
A second protocol, Agent2Agent (A2A), first developed by Google and also now under the Linux Foundation, lets agents from different vendors find each other and hand off work. MCP connects an agent to its tools; A2A connects agents to other agents.
Expect "Does it have an MCP server?" to join "Does it have an API?" on software buying checklists. A business app that supports these standards can be used by whichever AI assistant your team already pays for.
What changes at work first
The first jobs agents take over are back-office ones with many steps and quick checks: sorting support tickets, researching leads, assembling weekly reports, cleaning up records and making routine code changes. Gartner predicted in June 2025 that by 2028, 15% of day-to-day work decisions will be made autonomously through agentic AI, up from none in 2024.
A common setup pairs an agent that prepares the work with a person who approves anything that leaves the building, such as a sent email, a refund or a change to a customer record. For most roles, the near-term change is in the mix of tasks. Someone who spent mornings copying order details between two systems spends that time reviewing what an agent did and handling the cases it flagged as unusual.
Writing a process down before you hand it to an agent pays off, because an agent follows clear instructions well and guesses badly when the rules live only in one person's head.
Agents in creative work
In writing and marketing, agents mostly handle the research and assembly around the creative part: gathering sources, building outlines from the pages that already rank, drafting variations, making images and publishing. Wordze, for example, takes an article from keyword research through an outline and a draft to a Copyscape duplicate check and publishing on WordPress or Shopify inside one app.
The angle, the firsthand experience and the final judgment still come from a person, and so does the responsibility for what gets published. Our article on the ethics of AI article generation covers where the lines are.
What could slow agents down
Many projects will not last. Gartner predicted on June 25, 2025 that over 40% of agentic AI projects will be canceled by the end of 2027, because of rising costs, unclear business value or weak risk controls. It also warned about "agent washing," vendors relabeling existing chatbots, assistants and robotic process automation as agents, and estimated that only about 130 of the thousands of vendors claiming agent products offered the real thing.
Security is still an open problem. OWASP ranks prompt injection first (LLM01:2025) in its Top 10 for LLM applications: hidden instructions in a web page, email or file that the agent reads can change what it does. The more an agent can do, the more an attacker gains by tricking it. OWASP's advice is least-privilege access and human approval for privileged actions, which also limits how far agents can go on their own.
Accountability is also unclear. When an agent issues the wrong refund or emails the wrong customer, the business that deployed it answers for the mistake. Companies will want audit logs, spending caps and clear approval rules before they let agents act without a person watching.
Rules are arriving
The EU AI Act's transparency rules in Article 50 have applied since August 2, 2026. An AI system that talks directly with people must be designed so they know they are dealing with AI, unless that is obvious from the context. AI-generated text published to inform the public on matters of public interest must be disclosed as such, unless a person has reviewed it and takes editorial responsibility. The EU's Digital Omnibus regulation, in force since July 27, 2026, pushed the Act's requirements for high-risk AI systems back to December 2027 but left the Article 50 date unchanged.
If you sell to customers in the EU, a support agent on your site needs to say it is an AI, and that is a sensible default everywhere else too.
How to prepare
- Write down two or three repetitive processes, step by step, with the rules a new hire would need.
- Check which of your core apps support MCP or have agent features already.
- Pilot one agent on one process with read-only access, then add actions one at a time.
- Require approval for anything that sends, pays, deletes or changes customer records.
- Track the time saved and the errors caught for a month before you expand.
- Label AI agents clearly wherever customers talk to them.
For the basics of how agents work, start with our introduction to AI agents. To compare tools for building them, see the top AI agent platforms.
