Definition
A personal agent is an AI agent that serves an individual across tasks and sessions. It uses that person's context and preferences to pursue goals within authority the person has delegated. The agent might manage a calendar while tracking an unfinished project, then help investigate a purchase. Each task draws on the continuing relationship with the user.
In this glossary, agent means a system that can choose actions, use capabilities, inspect results, and continue working. Personal describes whom it serves and how it carries context forward. A friendly voice or a remembered name alone does not meet that threshold.
Personal agents in the glossary
These personal agents have dedicated glossary entries:
Origin and attribution
The idea predates language models. Pattie Maes's July 1994 paper, Agents that Reduce Work and Information Overload, described interface agents that learn from users and help with computer tasks. Her 1997 debate with Ben Shneiderman explicitly discussed agents that know a user's interests and act on the user's behalf. These are documented antecedents; neither establishes a sole inventor or first use of the phrase personal agent.
OpenClaw is a current implementation of the personal-agent pattern, with a continuing assistant accessible through messaging channels and a workspace for instructions and retained context. Other implementations can use hosted services or different memory designs. Personal agent does not prescribe a product, model family, or deployment location.
Scope and disagreement
There is no uniform technical threshold for how much continuity or initiative a product must provide to call itself a personal agent. This working definition requires an action loop serving an individual and continuity across tasks. A tool-free companion chatbot may offer a personal relationship without meeting that operational definition.
Maes argued for delegation to agents that understand user interests. Shneiderman favored predictable interfaces and reversible, visible actions; he questioned whether software could reliably infer intentions from vague goals. Their disagreement concerns how much control to delegate and how a person can understand the result.
Operational significance
Separate information the agent may read from actions it may take. Permission to inspect a calendar does not authorize sending invitations, and knowledge of a person's habits does not establish permission to spend money. Persist the user's corrections with provenance and a deletion path. Define where approval is needed and how the person can stop work or inspect changes.
Personal agents often encounter private messages and shared conversations. Context belonging to one person must not silently become context for everyone who can contact the agent. OpenClaw's documentation assumes a single trust boundary per Gateway. Check how the assistant separates access before connecting other users.
Distinguish it from nearby terms
- An AI agent is the broader action-loop category. A personal agent adds an ongoing service relationship with an individual.
- Agent memory supplies retained context. It can exist in a project agent or a shared service without creating a personal agent.
- A chatbot can provide answers and companionship while leaving every external action to the user.
- SOUL.md configures behavior and persona in compatible harnesses. It does not supply the agent's memory system or delegated permissions.
- Autonomy measures how independently a system can act. A personal agent can have narrow authority and frequent human approval.
Check your understanding
An assistant remembers your preferred meeting times and can draft calendar changes. What would make it a personal agent under this definition, and what separate evidence would establish permission to send an invitation to somebody else?