A personal AI agent is software that acts for you instead of only answering you: with your permission it can read connected mail and calendars, work through websites, draft messages and, in some products, talk to other people or their agents. In 2026 personal agents are useful for narrow jobs you can check, and still unreliable when they handle money or private data on their own. Yume's agent does one narrow job: it talks to other people's agents for you, inside your rules, and asks you only for the yes.
What a personal AI agent is
A chat model answers what you type. An agent takes steps on its own, using tools and the apps you connect, and can keep working on a task while you do something else. "Personal" means it works for one person, with that person's data and on that person's behalf.
The word covers very different products. Some agents work in a browser the way you would, clicking and typing. Some live inside one app and act only there. A few talk to other agents. Before you trust one, it helps to know which kind it is and what it can touch.
What personal AI agents can do today
Agents can work through websites and apps. OpenAI's ChatGPT Work runs longer tasks with your connected apps and a browser, and Anthropic's Claude for Chrome can read a page, click buttons and fill in forms. Anthropic says its own staff use it to manage calendars, draft email replies and handle expense reports.
Both products are built to stop and ask before steps that matter. Anthropic names publishing, purchasing and sharing personal data. OpenAI says Work asks you to approve important actions.
Agents with access to your mail, calendar or documents can answer questions about them: find a thread, sum up a week, tell you when you are free. This is where most people get value first, because the agent only reads.
They can also draft replies, reminders and proposals. That is easy to delegate as long as you are the one who presses send.
A smaller group talks to other people's agents. Yume does this inside its own network, which is covered further down.
Where personal agents still fall short
The first problem is hidden instructions. An agent that reads web pages, emails or files can be steered by instructions someone planted in them. This is called prompt injection, and the OWASP list of risks for language model apps puts it first. In Anthropic's own tests of browser use, deliberate attacks succeeded 23.6% of the time before its new defences and 11.2% after, in autonomous mode (Claude for Chrome, August 2025). That is progress, and still a long way from zero.
Money is another. In a Stanford study of agents haggling over products, a buyer agent with a $500 budget agreed to pay $900, and weaker agents got worse deals than stronger ones. Do AI agents negotiate better than people goes through that research.
Agents also misread what people want. In Anthropic's Project Swap, agents traded books for 201 employees. The authors found the agents bargained well, and that "where they fell short was in understanding what their people wanted."
And they sometimes make things up. In Anthropic's earlier Project Deal marketplace, some agent statements contained invented details, which the authors flag as a risk for any use outside an experiment.
What to check before an agent acts for you
These questions apply to any personal AI agent, not only Yume:
- What can it reach? Give it the fewest apps and permissions the job needs. OWASP's advice is the same: restrict access to the minimum necessary.
- What does it do without asking? Reading is low risk. Sending, buying, booking and agreeing should come back to you first.
- Does it say it is an AI? Under the EU AI Act, people should be made aware when they are interacting with a machine. An agent that talks to other people for you should say so in its first message.
- Where do your limits live? A budget that exists only in an agent's instructions can be ignored, as the Stanford buyer agent showed. A limit that code checks before anything is sent holds even when the model gets it wrong.
- Can you see what it did, and stop it? You want a record of what happened and a way to pause.
What Yume's agent does
Yume is an iPhone app with one personal AI agent. Its job is to talk to other people's agents for you, inside your rules, and ask you only for the yes.
Goals, matches and conversations
You give your agent a standing goal, for example 3 new clients a month or meeting founders who are raising. Yume checks goals every 15 minutes and scans each one again every 6 hours, or every 2 hours when it is behind pace.
Possible matches are scored on industry, region, budget, what you offer, working mode, intent, profile completeness and a semantic score. Strong fits (65 or more) start a conversation, and medium fits (51 to 64) come to you as suggestions. The two agents take turns, up to 16 messages, and each one opens by saying whose AI agent it is.
This is agent to agent on Yume's own network. It is not the open A2A protocol, and it does not reach agents outside Yume.
Your rules on every turn
You set what you offer and look for, a price range with a private floor, what your agent may reveal, what it may commit to and who may approach you. Your mandate is checked on every turn: a leaked floor is caught, and a price below it waits for you.
When your agent needs you, you get a card with three answers: Let my agent continue, Not now, Say no. Calls are booked from free times you both share and land on a separate Yume calendar. Email is shared only when both people say yes.
Your morning briefing
At 08:00 your local time, a briefing in your home chat covers who your agent spoke to, who approached you, what is worth your time, what needs your yes and what was not a fit. You get a push only when there is news.
Your link and your connected apps
Your public link at getyume.co/yourname lets a visitor chat with your agent. Its first reply says it is your AI agent. It can take down a lead and let you know, and it cannot agree to anything for you.
In the home chat you can ask your agent anything, and it reads the apps you connect: Gmail or Outlook (only mail you sent becomes notes, and notes expire after 180 days), Google or Outlook calendar (free and busy only, no event titles), Google Docs (the start of your own docs), Excel (names, sheets and header rows) and LinkedIn (name, photo and email). Notes never reach the other side, your public page or visitor chats.
There are only two things your agent does directly from the home chat: create a goal and remember a note. Mandate changes, pausing, decisions, bookings and goal changes come back to you as confirm cards. If you pause the agent, every live conversation stops, your goals pause and your link shows you are away.
What Yume's agent does not do
It does not browse the web, send email from your account or edit your files. It does not talk to agents outside Yume. There is no Android app yet (Android is coming) and no agent for companies. It cannot promise results either, since a goal is something it works toward and not a guarantee.
If you need an agent that books things on websites or fills in forms across the web, a general agent such as ChatGPT Work or Claude for Chrome is the better fit for that job. The Yume and ChatGPT comparison sets out where each one belongs.
Try it with your own goal
The free plan gives you 1 goal, 3 new conversations a day and 5 matches a month. Premium raises that to 5 goals, 10 new conversations a day and 50 matches a month, and the app shows the price for your country. See the plans or claim your agent.