We use cookies to understand how you use our site and improve your experience. Cookie policy
AI and agent glossary for teams | Palette
AI and agent glossary for teams
Last updated:
Plain-English definitions for teams working with AI and agents. This glossary covers common AI concepts, ways of working across a team, and the language used in Palette.
The focus is practical. Each definition explains what a term changes when people use AI at work. If a technical detail does not help you make a better decision, it is left out.
The software layer that runs an agent. It gives a model its instructions, tools, permissions, and working loop, then manages what happens between each model response. Claude Code, Codex, Gemini CLI, Mistral Vibe, and the Palette agent are different harnesses available in Palette Desktop.
Information an agent can retain and use beyond the immediate message. Memory can come from saved conversation history, files, instruction files, or a shared system such as the Context Library. It is different from a model's context window, which only limits how much information can be considered at one time.
A folder or shared environment organized so an agent can understand and work in it without a long explanation every time. Clear names, useful README files, written instructions, examples, and current decisions all help. Agent-ready does not mean machine-only. The same structure should make the work easier for people to navigate.
The person responsible for helping AI work across a company or team, regardless of their formal job title. They test tools, help colleagues adopt new ways of working, set practical rules, and often discover the company's context and governance gaps first.
A system that uses an AI model to work toward a goal over one or more steps. An agent can inspect information, choose tools, take actions, check the result, and continue. A model generates responses, while an agent adds the working loop and the ability to act.
A team that treats AI as part of how work gets done, not as an occasional writing assistant. People still set direction and make important decisions, while agents help research, draft, analyze, and execute. Buying AI licenses alone does not make a team AI-native.
A secret credential that lets software use a service, such as a model provider, on your account. When an agent uses your API key, usage is normally billed by that provider. Treat an API key like a password and never put it in a shared document, prompt, or public file.
A file or output produced by a person or agent that can be kept, shared, or used in later work. In Palette, artifacts include shared Markdown and interactive HTML outputs. The useful distinction is between the conversation where work happened and the artifact worth carrying forward.
A conversation with one agent inside a Palette Desktop session. A session can contain several chats working on different parts of the same task, while all of them share the session's working copy of the folder.
A saved snapshot of work in a Palette Desktop folder or session. Checkpoints create a history you can inspect and restore, so experimentation does not depend on remembering every change or undoing it by hand.
An organization-level link between Palette and a team tool such as Slack, Linear, GitHub, or Notion. Connections provide source material that can help keep the Context Library current. They are different from personal tools made available to agents through Connector Gateway.
An early Palette capability for connecting a personal tool once and making it available across the agents you use. The connection uses your own account and permissions. Google Calendar is the current pilot, so Connector Gateway should not yet be understood as a broad catalog of personal connectors.
The information available to a model when it produces a response. Context can include your prompt, conversation history, instructions, files, retrieved facts, and tool results. Better context usually improves relevance, but more context is not automatically better if it is stale, noisy, or contradictory.
The practice of deciding what information an AI should receive, how it should be structured, and how it stays current. Prompt engineering improves the request. Context engineering improves the information environment around the request.
The part of Palette OS that holds your company's shared, living docs for people and agents. A team can write context directly, link to material kept elsewhere, and let Palette maintain selected parts from connected work. The team controls what is included and what stays exactly as written.
An individual document in the Context Library, such as a company overview, team brief, or person profile. A page belongs to an organization, team, or user scope, which controls where it is available. Context pages are designed to be readable by both people and agents.
The maximum amount of information a model can consider in one request, measured in tokens. The window may contain instructions, chat history, files, and tool results as well as your latest message. A larger window helps with bigger tasks, but does not give a model permanent memory.
A repeatable way to test how well a model, agent, prompt, or workflow performs a defined job. An evaluation might check factual accuracy, whether instructions were followed, or whether an action completed safely. Good evaluations use realistic examples and clear criteria instead of judging a few impressive outputs.
AI systems that create new content such as text, images, audio, video, or code in response to instructions and context. The output is generated from patterns learned during training, which is why it can be useful and fluent without being guaranteed correct.
The rules and responsibilities that shape how a company uses AI. This includes approved tools, data access, retention, permissions, human review, and who is accountable when an agent takes an action. Good governance makes safe use easier instead of reducing everything to a ban.
Connecting an AI response to relevant sources instead of relying only on what the model learned during training. Files, company context, search results, and tool data can all provide grounding. Grounding reduces unsupported answers, but a source can still be incomplete or wrong, so important claims may still need review.
An answer generated by an AI model that sounds plausible but is unsupported or incorrect. A hallucination is not deliberate deception. The model is producing likely text rather than checking truth by default. Grounding, evaluations, and human review reduce the risk.
A package of work and context passed from one person or agent to another so they can continue without starting again. In Palette, a handoff can include a summary, instructions, references, and the intended recipient or team. It can be opened into a new or existing Desktop session.
A way of working where a person reviews, approves, corrects, or redirects an agent at meaningful points. The human does not need to approve every small step. The goal is to place judgment before consequential actions such as publishing, sending, spending, deleting, or saving changes into shared work.
A file that tells an agent how to work in a folder or project. Files such as AGENTS.md and CLAUDE.md can describe the team, structure, conventions, and checks an agent should follow. Different agent tools read different filenames, so the instructions should remain consistent across the files your team uses.
A model trained on large amounts of text so it can understand and generate language. Models such as Claude, GPT, Gemini, and Mistral can summarize, draft, classify, and reason over text. An LLM is the engine, not the complete agent that plans work and uses tools.
A plain-text format that adds simple structure with headings, lists, links, and other marks. Markdown files use the .md extension and remain readable without special software. That makes them useful for shared context because people and agents can both read and edit them reliably.
The AI engine that interprets input and generates output. Models vary in capability, speed, cost, context window, and whether they can work with text, images, or other media. The same model can be used through different agents, and one agent may let you choose between several models.
An open standard that lets AI agents connect to external tools and information through a common interface. An MCP server can offer resources to read and tools to call. Palette uses MCP to make company context, skills, handoffs, and supported connector tools available to agents.
A company or service that develops or operates AI models. Anthropic, OpenAI, Google, and Mistral are model providers. Your choice of provider affects available models, billing, data handling, and access methods, but it does not by itself determine which agent interface you use.
A model that runs on your own computer instead of sending each model request to a hosted provider. This can support offline work and keeps model inference local, but performance and model size depend on the computer's available memory and processing power.
The shared information needed to understand how a company currently works. It includes goals, responsibilities, decisions, customers, projects, constraints, and the reasons behind them. Tool data shows individual events, while organizational context explains what those events mean together.
The durable record of what a company has learned, decided, and changed over time. It can live in docs, logs, decisions, artifacts, and handoffs. Context helps with the work happening now, while memory helps a team understand how it got there and avoid repeating old work.
Palette's own agent harness inside Palette Desktop. It can use a model through Palette Cloud, run an on-device model, or connect to a compatible self-hosted endpoint. It appears beside Claude Code, Codex, Gemini CLI, and Mistral Vibe when you choose an agent for a chat.
A desktop app for running AI agents on the files and folders where your team works. A folder becomes a workspace, sessions create safe working copies, and you review changes before saving them back. Desktop supports the Palette agent, Claude Code, Codex, Gemini CLI, and Mistral Vibe.
The platform layer that holds your company's context, shared skills, and the work passed between people and agents. Its contents are scoped across the organization, teams, and users, then made available to supported agents through MCP. Palette Desktop is one way to work with the OS, not the only one.
Rules that determine what an agent can read, change, run, or send. Permissions may apply to folders, tools, accounts, or individual actions. Give an agent enough access to complete the task, then require review for actions that are hard to reverse or affect other people.
A Palette Desktop chat setting where the agent proposes an approach before it starts changing files. Plan mode is useful when the task is broad, the direction is uncertain, or you want to agree on the outcome before work begins.
The request or instruction given to a model or agent. A useful prompt makes the goal, relevant constraints, and expected result clear. It does not need to contain every piece of background if the agent can read that context from files, tools, or a shared library.
The practice of improving instructions so a model or agent produces a more useful result. This can include a clear role, examples, constraints, output format, and success criteria. Prompt engineering shapes the request, while context engineering shapes the information available around it.
An attack or failure mode where untrusted content contains instructions intended to redirect an agent. The instructions may be hidden in a webpage, email, file, or tool result. Permissions, trusted sources, and human confirmation before consequential actions help limit the damage.
A model optimized to spend more computation on multi-step problems before answering. Reasoning models can be useful for planning, analysis, and difficult tradeoffs, but they may take longer and use more tokens than faster general-purpose models.
A folder outside the current Palette Desktop workspace that you make available to an agent as additional context. The reference remains separate from the workspace, so you do not need to copy its contents into the folder you are working in. References belong to the person's Desktop setup rather than traveling automatically with a shared workspace.
A method where a system retrieves relevant information before asking a model to generate an answer. Retrieval can ground the answer in current or private sources that were not part of the model's training. RAG is one way to provide context, not a guarantee that the right source was found or interpreted correctly.
An isolated working area that limits where an agent can act. In Palette Desktop, a session gives work its own copy of the workspace, while chat permissions control whether the agent can reach beyond the folders you have made available. Sandboxing makes experimentation safer, but it does not replace reviewing consequential changes.
The level at which context, skills, or other shared assets are available. Palette uses organization, team, and user scopes so information can be shared broadly or kept close to the people who need it. Scope is about access and relevance, not just where a file happens to be stored.
A safe working copy of a Palette Desktop workspace. A session can hold several chats, lets agents change files without immediately changing the shared folder, and gives you a review step before saving work back. Several sessions can explore different directions in parallel.
Information that a team deliberately makes available across people and agents. Shared context can include strategy, decisions, roles, customer knowledge, and working conventions. It should be scoped and maintained, not pasted as one enormous prompt into every conversation.
A reusable package that teaches an agent how to perform a particular kind of work. A skill can contain instructions, examples, templates, reference material, and scripts. In Palette, built-in and custom skills can be shared at organization, team, or user scope.
Starting instructions supplied to a model by the application or agent harness. A system prompt can define behavior, boundaries, and how tools should be used before the user's prompt is added. It is only one layer of instruction, files and live context may add more.
A small unit of text processed by a language model. A token is not the same as a word, and the exact split depends on the model and language. Input and output tokens affect context limits, response time, and API usage costs.
The ability of a model or agent to request an action outside text generation, such as searching the web, reading a file, querying a system, or updating a document. The surrounding agent or application executes the tool and returns the result to the model. Permissions determine which tools and actions are allowed.
A repeatable sequence of steps that turns an input into an outcome. An AI workflow may combine people, agents, skills, tools, and approval points. The best workflow does not automate every step. It puts automation where it removes routine work and keeps human judgment where it matters.
A folder opened in Palette Desktop as the main place for a piece of work. The folder and its files are the source of truth, and it can be local or shared through a service such as Google Drive or Dropbox. Sessions let agents work on safe copies before changes are saved back to the workspace.
