--- title: "What's a Personal Agent? The AI That Keeps Working · Comma Blog" description: "A personal agent keeps context, carries work across time, and returns control when judgement is needed. Learn what separates one from a chatbot and try Comma." url: https://comma.surf/en-gb/blog/what-is-a-personal-agent language: en-GB --- [Back to blog](https://comma.surf/en-gb/blog) Comma Team · 26 October 2026 · 10 min read # What's a Personal Agent? The AI That Keeps Working A personal agent keeps context, carries work across time, and returns control when judgement is needed. Learn what separates one from a chatbot and try Comma. [Get started](https://app.comma.surf/) [Read the docs](https://comma.surf/en-gb/docs/agent) A personal agent is an AI system that works for one person across more than one conversation. It keeps the context needed to continue a responsibility, uses tools or computers to make progress, and brings decisions back to the person when judgement or permission is required. That definition is more demanding than “a chatbot that knows your name.” A chatbot can produce an excellent answer and still leave you responsible for carrying the work between messages. A personal agent has to preserve a useful state: what outcome you asked for, which materials matter, what has already happened, what remains blocked, and when it should involve you again. The most revealing question is therefore not *How intelligent is the model?* It is *What happens after I leave?* If the work disappears with the session, the system may be helpful, but it is not yet acting as a continuing agent on your behalf. ## A practical definition of a personal agent The broader category of an AI agent is usually defined by its ability to pursue a goal and take actions through tools. [Google Cloud's overview of AI agents](https://cloud.google.com/discover/what-are-ai-agents), for example, distinguishes agents from systems that only generate a response by emphasising reasoning, planning and tool use. “Personal” narrows that relationship: the agent serves an individual rather than an organisation’s queue, a public workflow or a single application. In practical terms, a personal agent combines five responsibilities: 1. **It works toward your outcome.** The assignment belongs to a person, not merely to a temporary prompt. 2. **It maintains working context.** It can continue a Task without asking you to reconstruct the request at every step. 3. **It can act in an environment.** Depending on the product and the permissions you grant, that may include connected apps, files, a browser, a local computer or a cloud computer. 4. **It persists across time.** It can wait, resume and react to relevant events instead of requiring an open chat window. 5. **It returns control at meaningful boundaries.** A good agent distinguishes preparation it can continue from decisions, permissions and consequences that belong to you. None of these properties requires the agent to imitate a person. It does not need a face, a human biography or a fictional personality. The relationship is defined by continuity and responsibility: one system is carrying work for one user, within boundaries that user can inspect and change. A personal agent is a loop around an ongoing responsibility, not a single model response. It makes progress where it has access and returns to the user for review, permission or a final result. ## The session is the wrong boundary Most conversational products are organised around sessions. You open a thread, provide context, receive an answer and decide what to do next. That shape works well for questions, brainstorming and bounded transformations. It becomes awkward when the request contains waiting, dependencies or several places where work must happen. Imagine asking for a weekly project brief. Producing it may require checking a repository, reading a planning document, noticing a delayed milestone and waiting for a late status update. A response-only system can tell you how to assemble the brief. A personal agent needs somewhere for the responsibility to persist while inputs arrive and work progresses. This is why “sessionless” is a useful test. It does not mean that conversations have no boundaries or that an agent remembers everything forever. It means the life of the work is not constrained to the life of one chat session. The Task can remain in progress, wait for an event, ask for a decision and resume from the same working state. Our guide to how an always-on personal agent keeps long-term work moving examines that operating state in detail. That continuity also changes the user’s role. You should not have to keep the system alive by sending “continue” or repeatedly asking whether anything changed. Your responsibility shifts toward setting the outcome, supplying access, answering genuine decisions and evaluating the result. ## Always-on does not mean always acting An always-on personal agent can keep watch and continue suitable work while you are elsewhere. It should not interpret persistence as permanent permission to do anything it can reach. The useful form of always-on operation has several modes. Sometimes the agent is actively working. Sometimes it is waiting for a file, a message or a scheduled time. Sometimes a lightweight process is screening events so the main agent does not repeatedly wake up just to discover that nothing matters. And sometimes the Task is deliberately paused because the next choice belongs to you. This distinction matters both for attention and cost. Polling every connected source with a large model would waste resources and produce unnecessary decisions. Comma’s [programmable event-loop architecture](https://comma.surf/en-gb/blog/agent-programmable-event-loops) uses small Loops to watch events and wake the main agent only when something deserves further work. The technical implementation is specific to Comma, but the product principle is broader: a persistent agent should be quiet by default and active for a reason. A separate guide explains when an always-on personal agent should wake up by connecting relevance, timing and authority to the next response. Always-on should therefore describe availability and continuity, not unchecked autonomy. A credible product should make it possible to tell whether work is running, waiting, blocked, ready for review or finished. ## A personal agent needs somewhere to work Reasoning alone does not update a spreadsheet, compare files, navigate a website or prepare an artifact. An agent needs an execution environment, and different environments create different capabilities and responsibilities. | Environment | What it enables | What still limits it | | --- | --- | --- | | Connected apps and APIs | Structured access to mail, calendars, documents and project systems | The available connector, granted scope and action permissions | | A browser or cloud computer | Work that spans websites or needs a persistent remote environment | Authentication, site behaviour, available files, cost and approval rules | | Your local computer | Access to local files and applications | Device availability and explicit local permissions | | A self-hosted server | Control over the application stack, databases and model configuration | Your infrastructure, maintenance, backups and external providers | No single environment is automatically best. Cloud execution can continue after a laptop closes, but it cannot magically access a file that exists only on that laptop. Local access can reach personal material, but the device and its permissions become part of the task. Self-hosting increases control over the stack and stored data while making the operator responsible for keeping it healthy. The location of execution is not a minor implementation detail. It determines what the agent can reach, when it can work, where state is stored and who is responsible when something fails. ## Memory should serve the work, not become a slogan A personal agent needs memory, but “memory” covers several different things. There is conversation history, durable information about the user, and the working state of a specific Task. Treating them as one unlimited store creates confusing expectations. For an ongoing Task, the agent may need the desired outcome, the relevant files, decisions already made and the current blocker. That is working memory for a responsibility. A long-term user preference, such as a preferred report format, may be reusable across Tasks. A one-time password or an obsolete draft should not become permanent personal knowledge merely because it appeared in a conversation. More memory is not automatically more personal. Useful memory is relevant, correctable and governed. The user should understand which context is affecting the work and be able to correct a mistaken assumption. Persistence without correction turns yesterday’s error into tomorrow’s action. Comma describes one continuing agent carrying work, memory and context across devices and conversations. That promise should be read with the same practical boundaries as any connected system: context depends on what has been provided or authorised, and an agent can still misunderstand incomplete or conflicting material. ## The handoff back to the person is part of the system Personal agents operate in a world where many choices cannot be inferred safely from available data. A document can contain contradictory priorities. An email can be technically ready to send but politically unwise. A form can be complete while its legal declaration still belongs to the person named on it. A well-designed agent does not hide these boundaries. It prepares what it can, identifies the decision and returns the Task with enough context for the user to act. The review step is not a failure of autonomy. It is part of making delegation usable. Comma represents this explicitly with Task states including **In progress**, **Needs Review** and **Done**. Needs Review is where an ambiguity, permission or consequential decision comes back to the user. Done means the agent has completed its current work and delivered an output; it does not certify that every fact or judgement is correct. This creates a more honest division of responsibility. The agent owns progress within the assignment. The person owns the assignment’s meaning and consequences. ## Hosted and self-hosted are two ownership choices Deployment determines who operates the system; the hosted versus self-hosted personal-agent guide explains the resulting control, data and maintenance responsibilities. ## What a personal agent can do well The best early assignments have a clear outcome, require continuity, and produce something the user can inspect. Preparing a brief from several files is a better fit than “make good decisions for me.” Monitoring a bounded event and drafting a response is more testable than “handle my communications.” Organising materials for review is safer than silently treating an ambiguous source as authoritative. Useful assignments often share three properties. The agent can recognise progress, the available environment contains the necessary inputs, and there is a clear point where the result returns to the user. Examples include preparing recurring reports, following a defined project thread, organising research, monitoring a repository for a specific change, or assembling a draft from connected documents. Poor assignments conceal the success condition or delegate authority the user has not actually granted. “Take care of everything” gives neither the agent nor the user a reliable way to know what done means. A personal agent becomes more useful when the assignment is concrete enough to carry and the boundaries are visible enough to trust. ## How to evaluate a personal agent Product pages will increasingly use similar words: memory, proactive, autonomous and always-on. Rather than treating those labels as capabilities by themselves, follow one realistic Task through the product. Ask where the Task lives after the initial conversation. Check whether you can see its current state and whether the agent can wait without losing the assignment. Identify which environment performs the work and what happens if that environment is unavailable. Look for a visible review boundary before consequential actions. Finally, ask what you can inspect, revoke, export or operate yourself. This test exposes the difference between a compelling demonstration and a system you can actually delegate to. Intelligence matters, but so do state, access, persistence and control. A personal agent is the product of all five. ## How Comma approaches the category [Comma](https://comma.surf/en-gb/) is a free, open-source personal agent built around one continuing relationship rather than a collection of disconnected sessions. Quick questions can stay in chat; larger requests become Tasks with visible progress. Routines prepare recurring views of what needs attention. Its own cloud computer can continue suitable work in the background, and connected Macs provide a separate local environment with read-only access by default until operations are allowed. The same product can be used as a hosted service or [run on infrastructure you manage](https://comma.surf/en-gb/self-host). Signal and Telegram provide remote ways to hand off work, while the Task remains the continuing unit behind the conversation. These pieces support Comma’s central claim: **your 24/7 personal agent, with no full stop until the work is done.** That does not mean every Task succeeds, every integration is available or every action happens without review. It means work has somewhere to continue, a state you can follow, and a defined way to return to you. The full product walkthrough in [Meet Comma](https://comma.surf/en-gb/blog/meet-comma-personal-agent) shows how those pieces fit together in an illustrative knowledge-work scenario. ## Questions about personal agents ### Is a personal agent just a chatbot with memory? No. Memory can improve a chatbot, but a personal agent also needs a continuing unit of work, access to an environment where it can act, and a way to wait, resume and return control. A long conversation may preserve text without preserving responsibility. ### Does every personal agent have to be always-on? Not every personal agent needs to monitor events continuously. Preserving a relationship and task state across conversations is different from remaining available to detect changes and resume in the background. What's an Always-On AI Agent? explains the additional system conditions for that background availability. ### Does a personal agent need its own computer? Not in every architecture. Tools and APIs can complete many Tasks. A browser, cloud computer or connected local device expands the work an agent can perform when an API is missing or files live in a specific environment. The important question is whether the execution environment matches the assignment. ### Is self-hosting automatically more private? No. Self-hosting changes who operates the application and databases; privacy still depends on the complete path through models, search services, connected apps and network infrastructure. See the hosted and self-hosted deployment comparison before treating deployment location as a privacy guarantee. ### Can a personal agent act without asking every time? That depends on the action, the permissions and the product’s controls. Reversible preparation can often continue within a bounded Task. External communication, destructive changes, purchases and decisions with meaningful consequences deserve explicit review or narrowly defined prior rules. ### What is the best first Task for a personal agent? Choose work with known inputs and an inspectable result: prepare a weekly brief from specified sources, monitor a defined change, or organise a set of files into a draft. Start with read-and-prepare work, evaluate the output, and expand permissions only when the workflow has earned your trust. ## Give continuing work a place to live A useful personal agent does not merely answer in your voice or remember a few preferences. It carries a bounded responsibility across time, works in environments you authorise, and knows when the next step belongs to you. Start Comma with one piece of work whose outcome you can recognise. Follow the Task rather than keeping a chat alive, review the moments that require your judgement, and decide which ownership model fits you. You can [get started with hosted Comma](https://comma.surf/en-gb/), read the [documentation](https://comma.surf/en-gb/docs), or choose the [self-hosted path](https://comma.surf/en-gb/self-host). Personal Agents [Back to blog](https://comma.surf/en-gb/blog) Share ## Related posts 20 October 2026 ### [Meet Comma: the personal agent that carries work to done](https://comma.surf/en-gb/blog/meet-comma-personal-agent) 19 October 2026 ### [Native SSH for Comma agents](https://comma.surf/en-gb/blog/native-ssh) 1 October 2026 ### [Programmable event loops for our distributed agent harness](https://comma.surf/en-gb/blog/agent-programmable-event-loops) ## Try Comma, Now. [Get started](https://app.comma.surf/)