What AI agents are, and where they are going

Agent is the word of the year in our industry, and it is used for almost anything: a chatbot with a new name, a script that sends emails, a vision of companies that run themselves. That makes it hard to tell what is real.

Here is the plain version, with the evidence as it stands in October 2026.

What an agent is

An agent is a language model given a goal instead of a single question. It plans the steps, uses tools to take them, checks each result against the goal and keeps going until the work is done or it needs a person. Anthropic puts it in one line: a model that directs its own processes and tool use when accomplishing a task.

The difference to what came before is who decides the next step. A chatbot answers, and you decide what to do with the answer. An assistant does one thing you ask inside the conversation: a draft, a summary, a lookup. A workflow follows a path someone wrote in advance, with a model doing parts of it. An agent chooses its own path. Given the goal of preparing an offer, it reads the last emails, looks up the prices, drafts the document and asks before it grants a discount.

Agentic is the word for the whole family: systems in which a model acts, not only answers. A system is more or less agentic depending on how much of the path it chooses and how much it may do without asking. That is decided by the people who build and run it, not by the model.

How they work

Four parts turn a model into an agent. The first is tools: search, read a file, query a database, write to a calendar, send a message. Tools reach a company's systems through connectors, and the common standard for those is MCP, the Model Context Protocol. Anthropic introduced it in 2024 and in December 2025 handed it to the Agentic AI Foundation, a fund under the Linux Foundation that it founded with Block and OpenAI. By then there were more than 10'000 active public MCP servers, and ChatGPT, Gemini and Microsoft Copilot had adopted the protocol.

The second is memory: what the agent keeps between steps and between sessions, so it knows what it has done and how the company likes things done. The third is computer use, for systems without a connector. The agent looks at the screen, then clicks and types like a person. That is slower and more fragile than a connector. Claude Cowork, Anthropic's agent for office work, reaches for connectors first and takes over the screen only when it has to.

The fourth is the harness around the model: the instructions, the limits and the checks that decide what it may do. Larger jobs are split between several agents. An orchestrator plans, hands parts to worker agents that run in parallel and joins their results. In Anthropic's research system, that setup beat a single agent by 90.2 percent on an internal research test, and used about 15 times the tokens of a chat. It pays for broad research, much less for work where every step depends on the last one.

Where they stand in October 2026

Every large lab now ships agents as products. On 29 September, OpenAI launched Dots: agents in ChatGPT that each work from a cloud computer of their own and keep going towards a goal between conversations. Anthropic offers Claude Code for software and Claude Cowork for office work, which acts in the folders and tools a user chooses and keeps running in the cloud when the laptop is closed. On 8 October, Google Cloud introduced the Gemini agent, which plans the work, uses skills and tools and connects to a company's systems.

What works reliably is work with a clear end and a quick check. Software leads: in Anthropic's data from February, software engineering made up nearly half of the tool calls on its public API, while fields such as customer service, sales and finance each made up a few percent at most. Computer use has improved fast. On the verified version of OSWorld, a benchmark of everyday computer tasks, the best agent scored 85 percent in June 2026, against 42 percent a year before and roughly 72 percent for people on the same tasks. The same analysis, by Andreessen Horowitz, adds the other side: 85 percent still means 15 tasks in 100 failed, and a task a person finishes in two to three minutes can take an agent eight to ten.

What does not work yet is length, judgement and safety. METR, which rates agents by how long their tasks would take a person, warns that its headline number is no licence to delegate: it marks the tasks agents finish half the time, and some work needs 98 percent success before it is worth automating. Agents also stay brittle when the work drifts off the runbook. And prompt injection, instructions hidden in a page or an email that an agent reads and follows, is still open. OpenAI wrote in December 2025 that, like scams on the web, it is unlikely ever to be fully solved.

What it means for a company

Not every job needs an agent. Where the path is known, a workflow is cheaper and more predictable. An agent earns its cost where the path changes from case to case: research, an unusual support request, the preparation of an offer. Many good systems are mostly workflow, with an agent at the one or two steps that need judgement.

Where an agent does the work, decide what it may touch. Give it the narrowest access the job needs, and set every action to one of three states, the ones Anthropic offers in Claude: always allowed, needs approval or blocked. Put the approvals before anything that cannot be undone: the payment, the email that leaves the company, the published page. Build them into the tools, not only into the prompt. An instruction can be talked around. A missing permission cannot.

Then make the work traceable. Run every write under the name of the person who asked for it, and keep a record of every step: what the agent read, which tool it called, what it changed. Twin, the AI brain we build for companies, works this way: boundaries, approvals and rights by role, and every step logged.

Where it is going

The direction is steady. By METR's long run estimate, the length of the tasks agents can complete doubles every six to seven months. Its caution applies here as well: doubling the horizon does not double what can be handed over, because failures grow more complex and more costly.

The plumbing is settling too. MCP now sits in a neutral foundation backed by the large labs, so one connector can serve agents from different vendors, and a company that connects its systems once is not tied to one model.

Our line says it in six words: Human led. AI powered. Autonomous next. Now, people bring the idea, the experience and the taste, AI does the production, and people review and give the final touch. Next is agentic: agents do most of the work, and people set the direction and give the final yes. After that, all autonomous: people step in where they choose to. Every crue product moves along that line, one stage at a time.

Helm is where the line leads: our vision of a whole company run by its founder with agents. It is in the making. Even there, a person sets the direction and answers for the result. Autonomous will mean more work done without asking. It will not mean that nobody is responsible.