What is action space?
The set of actions an agent is allowed to choose from at a given step.
01 · Foundations · 103 terms
What belongs in foundations?
Foundations covers the basic objects and behaviors that make an agent an agent. This section has 103 definitions. Open any term for a plain answer, the source behind it, and the day we checked it. Related links take you to the ideas that explain it or show where it sits in the wider field.
103 shown
What is active RAG?
A search method where the system chooses what to find, checks it, and may search again before it answers.
What is agency?
The power to choose goal-led actions within a given setting.
What is agent card?
A file that tells other systems who an agent is, what it can do, and how to reach it.
What is agent discovery?
The process by which a user or another system finds an agent and learns what it can do.
What is agent loop?
The repeated cycle in which an agent observes, decides, acts, checks the result, and chooses what to do next.
What is agent manifest?
A file that lists an agent’s purpose, tools, limits, inputs, outputs, and rules.
What is agent payments?
Ways for agents to ask for, approve, send, or receive money within set limits.
What is agent registry?
A list of agents, what they can do, and where other systems can reach them.
What is agent skill?
A reusable set of instructions and tools that gives an agent a named ability.
What is agent team?
Two or more agents with assigned roles that cooperate on one goal.
What is agentic?
Able to seek results on its own, especially when used of AI.
What is agentic AI?
AI designed to choose and carry out actions toward a goal, often through one or more agents.
What is agentic commerce?
Trade in which agents help find, buy, sell, or pay for goods under set rules.
What is agentic RAG?
RAG in which an agent plans and changes its searches to meet a goal.
What is AI agent?
A system that observes a setting, chooses goal-led actions, and acts on that setting.
What is approval gate?
A point where an agent must stop and get consent before it takes a high-risk step.
What is artificial intelligence (AI)?
A machine system that predicts or makes choices for goals set by people.
What is autonomous tool use?
Tool use in which the agent chooses when, why, and how to call a tool without approval for each call.
What is autonomous validation?
A check that the system initiates and runs on its own before accepting or advancing work.
What is autonomy?
The degree to which an agent can choose and carry out actions without new human direction.
What is background agent?
An agent that works without an open user session. A clock or event may start it.
What is blackboard architecture?
A design in which several agents or modules read from and write to one shared working state.
What is capability discovery?
The process of finding what tools, skills, services, or agents are available for a task.
What is chain of thought?
Text or tokens a model produces as steps before an answer. They may not show its true internal reasoning.
What is circuit breaker?
A control that stops an agent after a risk alert, repeat failure, cost limit, or unsafe state.
What is claim grounding?
The practice of tying a factual statement to evidence that can be checked.
What is compounding error?
A mistake that changes later choices, causing the size or cost of the error to grow across a run.
What is confused deputy?
A failure where someone tricks a trusted system into misusing its access.
What is durable execution?
Work that can survive a restart, wait for an event, and resume from saved state.
What is dynamic model?
A system that selects or changes the model it uses according to the task, state, cost, or risk.
What is embodied environment?
A real or simulated world where an agent acts through a body, robot, or set of sensors.
What is environmental interaction?
The exchange in which an agent observes a world, acts on it, and receives new information in return.
What is error recovery?
Finding a failure, returning to a safe state, and choosing the next useful step.
What is escalation?
Sending a choice or problem to a person or system with more power or skill.
What is event-driven agent?
An agent that starts or changes work when a declared event occurs.
What is excessive agency?
A risk in which an agent has more functions, permissions, or freedom than its task needs.
What is feedback learning?
A way to improve later choices by using signals from past results.
What is freshness window?
The maximum age evidence may reach before it must be checked again.
What is function calling?
A model feature that names a function and gives the inputs for software to run it.
What is generality?
The range of tasks and settings in which a system can work without being rebuilt for each one.
What is goal-directedness?
How well a system keeps its choices aimed at a stated result.
What is graph of thoughts?
A method that maps ideas as a graph so they can branch, link, and change.
What is heuristic search?
Search guided by an estimate of which options are most likely to lead to a good result.
What is human in the loop?
A design in which a person must take part in selected decisions or stages of the work.
What is human on the loop?
A design in which an agent acts on its own while a person monitors and can intervene.
What is idempotency?
A property that lets the same action run again without causing a second change.
What is indirect prompt injection?
An attack where bad instructions hide in data an agent reads, such as a page, file, email, or tool result.
What is interrupt?
A signal that pauses or stops an agent so its state can be inspected, changed, or resumed.
What is judge model?
A model used to score or compare another model’s output or behavior.
What is large language model (LLM)?
An AI model trained on vast amounts of text to understand and generate language.
What is least privilege?
The rule that a user, agent, or service should receive only the access needed for its current task.
What is LLM as judge?
The use of a language model to grade, rank, or critique outputs under a stated rubric.
What is LLM-profiled role?
An agent role shaped mainly by instructions about its skills, duties, and limits.
What is long-running agent?
An agent designed to continue across long waits, many steps, or more than one user session.
What is Moltiverse?
A name used by the community around OpenClaw.
What is Monte Carlo tree search?
A plan search that samples possible outcomes and spends more effort on promising paths.
What is natural-language interaction environment?
A setting where an agent mainly sees and acts through written or spoken words.
What is pass at k?
The chance that at least one of k attempted outputs passes a test.
What is passive tool use?
Tool use in which software or a fixed workflow decides when the tool runs rather than the model deciding.
What is passive validation?
A check that observes or scores work without controlling what happens next.
What is perceive-act loop?
A cycle where an agent sees its setting, takes an action, and checks the result.
What is plan-and-execute?
An agent pattern that first creates a plan and then carries out its steps, revising when needed.
What is prompt injection?
An attack that puts commands in model input to override the task or rules.
What is provisional claim?
A claim shown with a warning because its evidence is not yet strong enough.
What is RAG-style tool use?
Tool use that finds outside facts and puts them into the model’s context.
What is ReAct?
An agent pattern that puts thought, action, and observation steps in a loop.
What is red teaming?
The deliberate search for ways a system can fail, be misused, or be attacked.
What is reflection?
A step in which an agent reviews earlier work and uses the review to change what it does next.
What is role-based agent?
An agent set up for a clear job, duty, or field of work.
What is router agent?
An agent that decides which specialist, model, tool, or workflow should receive a request.
What is rule-based environment?
A setting whose state or replies change through explicit rules.
What is sandbox?
A closed setting that limits what code or an agent can reach and change.
What is search graph?
A graph of possible states or ideas, linked by available moves or known ties.
What is search traversal?
The order and method used to visit options in a search tree or graph.
What is search tree?
A branching map of choices and the states that may follow them.
What is secret exfiltration?
Stealing passwords, keys, private data, or other guarded facts from a system.
What is self-critique?
A model’s review of its own work against clear rules.
What is self-refinement?
A loop where a model reviews and rewrites its own output.
What is service account?
An account that lets software or a workload prove its identity and gain access.
What is shared state?
Information that more than one agent or component can read or change during a task.
What is skill registry?
A list of reusable skills, what they do, how to call them, and what they need.
What is source provenance?
A record of where information came from and how it was collected or changed.
What is supervisor agent?
An agent that assigns work, watches other agents, and chooses when to retry, raise an issue, or stop.
What is task environment?
The world, tools, rules, and feedback within which an agent attempts a task.
What is task success?
A defined measure of whether an agent achieved the intended outcome.
What is termination condition?
A rule that tells an agent loop when to stop.
What is tool calling?
The process by which a model or agent requests that software run a named tool with stated arguments.
What is tool discovery?
Finding tools and learning their names, inputs, outputs, and limits.
What is tool error?
A failure returned when a tool cannot finish a request as asked.
What is tool invocation?
One specific request to run a tool with a set of arguments.
What is tool poisoning?
An attack that alters a tool, its description, or its output so an agent is misled or compromised.
What is tool registry?
A directory that lists tools and the information needed to discover and call them.
What is tool result?
The data or error returned after a tool invocation.
What is tool schema?
A structured description of a tool’s name, purpose, inputs, and output shape.
What is tool selection?
The decision about which available tool, if any, should be used for the current step.
What is tool use?
The use of an outside tool for its result or for the change it causes.
What is trajectory?
The ordered record of an agent’s states, actions, inputs, and outputs during a run.
What is tree of thoughts?
A method that explores several branches of thought and compares their value.
What is verifiable execution?
Work that leaves enough evidence for another party to check what ran and which rules it followed.
What is web conduct?
The rules and expected behavior that govern how an agent browses, collects, and acts on the web.
What is worker agent?
An agent that does tasks assigned by a lead agent or system.
What is world model?
A map an agent uses to predict how its world may change after an action.