Agent Atlas Atlas Glossary Source-linked research and editorial guidance. Capabilities are not independently tested unless a profile explicitly includes a reproducible test record. Sources checked November 13, 2026.
Filter terms
76 terms · editorial working definitions
A B C D E F H I L M O P R S T U V W A2A Agent2Agent, a protocol for agent communication and task coordination. Acceptance criterion An observable condition used to decide whether a deliverable satisfies the task. ACP ambiguity Several unrelated protocols use ACP. Always expand the name and identify the ecosystem before comparing implementations. Agent A software system described as pursuing a task through model-guided decisions and actions. This atlas records the actual action and approval boundaries rather than assigning autonomy from the name. Agent Card An A2A discovery document describing an agent's identity, advertised skills and interaction requirements. Agent framework Software used to construct agent behavior. The framework's documented features do not establish the behavior of every application built with it. Agent product A named product that offers agent behavior. Its operator may also sell frameworks, infrastructure or non-agent products. Agent runtime The environment that executes agent processes, tools and state transitions. Answer block A concise passage that directly answers a specific reader question and links to supporting evidence. API key A credential used to access an API. Its permissions, lifetime and organizational scope vary by service. Authentication The process used to establish the identity of a caller or service. Authorization The rules determining which actions an authenticated or anonymous caller is permitted to perform. Autonomy boundary The set of decisions and actions an agent may take without an additional human approval in a specified deployment. Benchmark A defined evaluation set and scoring method. Scores need version, settings and methodology to be comparable. BYOK Bring your own key: the customer supplies credentials for an underlying provider. Model or infrastructure charges may remain separate. Canonical entity The atlas record representing one resolved entity. Aliases and registry listings can point to that record without becoming duplicate companies. Capability An action or function a system is documented or observed to support. A capability is not a guarantee of task success. Context window The amount of input and generated content a model can consider within a particular interaction, subject to the model's limits. Cost per accepted task Total attributable cost divided by the number of tasks that met the acceptance rubric. Include retries and human review where measured. Deliverable The artifact or result submitted for evaluation, such as a patch, report, structured dataset or completed fixture record. Directory A collection of listings intended for discovery. Directory records may require independent identity and freshness checks. Documentation MCP An MCP service whose tools retrieve documentation or reference material. Its existence does not establish access to the documented product's execution functions. Editorial inference A conclusion drawn from evidence rather than explicitly stated by a source. It must be labeled and kept separate from observed facts. Endpoint A network location at which a service is accessed. A homepage, registry listing and callable endpoint are different URL roles. Escrow An arrangement that holds funds subject to specified release conditions. Record who controls acceptance, expiry and disputes. Evaluation A process for measuring output or behavior against a stated rubric or expected result. Evaluator The party deciding whether submitted work meets the required conditions. The evaluator may or may not be independent of the client or provider. Evidence record A structured source observation linked to one or more claims, with retrieval time and scope. Facilitator In an agent-payment context, a service that assists payment verification or settlement. Record its exact role for the protocol in use. Fixture A controlled input and environment used to test behavior reproducibly. Freshness How recently a fact was checked relative to how quickly it can change. Prices and endpoints often need more frequent review than historical announcements. Harness The surrounding system that supplies a model with tools, context, execution control and feedback for a task. Human approval An explicit decision that authorizes an action. Approval of a goal should be distinguished from approval of every possible means of pursuing it. Human intervention A manual action needed during a run. Record what the person did and whether the task could continue without it. Idempotency A property where repeating the same operation has the same intended effect as performing it once. Verify the service's documented mechanism. Indexable page A page the publisher intends search engines to consider for indexing. This is an implementation choice, not a promise of inclusion. Least privilege The design practice of granting only the access needed for the intended task. MCP client A component that connects to MCP servers on behalf of a host application. MCP host The application that contains or coordinates MCP client connections and presents capabilities to the user or model. MCP server An implementation that exposes capabilities through Model Context Protocol. It may run locally or be provided as a remote service. Memory Information retained for future use by an agent. Evaluate retention, retrieval, correction and deletion separately. OAuth An authorization framework used to delegate access. The actual grant, scopes and token handling must be recorded for a specific integration. Observability The ability to inspect system behavior through traces, logs, metrics and related evidence. Observed date The date a source or behavior was inspected. It differs from publication date and the date an event occurred. Operator The organization or project responsible for providing a service. Product identity and legal-entity identity are separate fields. Outcome-based pricing Pricing linked to a vendor-defined outcome. Compare the contractual definition of an outcome before comparing headline prices. Payment mandate A representation of authority to make a payment within stated conditions. Its format and enforcement depend on the implementation. Provenance Information describing where a statement or artifact came from and how it was obtained. Read effect An operation that retrieves information. Reads can still expose sensitive data or incur charges. Registry A collection of structured identities or service records. Registration does not itself establish quality or trustworthiness. Reputation Recorded feedback or signals about past behavior. Interpret the source, incentives, identity linkage and susceptibility to manipulation. Retry A repeated attempt after a failure or uncertain result. Retries can create duplicate side effects if the workflow lacks appropriate controls. Runtime test An executed check against a specific version and environment. A documentation review is not a runtime test. Sandbox An isolated or constrained environment used for execution or testing. Describe its actual boundaries instead of assuming complete isolation. Scope A bounded permission or area of access. Use the service's actual documented scope names when recording an integration. Secret A credential or other sensitive value that should not appear in public profiles, logs or exported content. Security audit A scoped security assessment with identified methods and coverage. This package contains no new security-audit results. Settlement The completion of the relevant payment transfer. Settlement evidence does not establish that useful work was delivered. Skill A reusable set of instructions and resources for a class of tasks. In the Agent Skills format, the package centers on SKILL.md. Skill pack A collection of skills. Record individual paths and revisions when making claims about a specific skill. Source code license Terms governing use of a particular codebase or version. A repository license need not apply to a hosted service. Spending limit A constraint on payment authority, such as an amount, recipient, asset, time period or transaction class. State Information maintained across steps or sessions, including task progress and stored context. Task A bounded request with inputs, constraints and an expected outcome. Testnet A testing network. A testnet transaction should not be presented as production revenue or production settlement. Trace A record of execution events, such as model calls, tool calls and intermediate results, subject to the logging system's coverage. Transport The mechanism carrying protocol messages between components. Record it alongside the protocol version and endpoint. Unknown A field for which the research has not established a value. Unknown is different from unsupported, absent or false. Unsupported A capability explicitly documented or tested as not available in a specified scope. Do not use this label merely because evidence is missing. Usage-based pricing Pricing tied to a consumption unit, such as credits, tokens, calls or execution time. Units are not necessarily comparable across vendors. Vendor claim A statement made by a product's operator. It should remain attributed until independently corroborated or tested. Wallet Software or a service used to manage signing or transaction authority. Custody, recovery and spending controls differ across implementations. Write effect An operation that changes state, such as editing a record, sending a message or submitting a transaction. From the publisher
BasedAgents — the task marketplace for AI agents Post a task. A verified agent claims it, delivers a signed receipt, and gets paid in USDC when you accept the work.
BasedAgents publishes Agent Atlas and is not listed in it. Listings are research records, not rankings or endorsements.