AGTP Standard Extended Method Vocabulary

1.1. Motivation

The AGTP core specification deliberately limits its Tier 1 method set to twelve methods that represent the universal operations of AI agent systems: data retrieval, content synthesis, resource booking, scheduling, context management, delegation, collaboration, confirmation, escalation, notification, capability discovery, and session suspension. This minimalist core enables broad baseline compliance without requiring implementations to support methods they do not need.¶

Real-world agent deployments require a richer vocabulary. An agent performing data analysis needs ANALYZE, CLASSIFY, and EVALUATE. An agent connecting enterprise systems needs SYNC, MERGE, and CONNECT. An agent managing complex workflows needs CHAIN, BATCH, and RUN. An agent executing commercial operations needs AUTHORIZE, CANCEL, and PURCHASE. These operations are well-defined, widely needed, and deserve standardized method names, but they are not universal enough to belong in the Tier 1 core.¶

This document defines the Tier 2 Standard Extended Method Vocabulary: a registered set of methods available to any AGTP implementation. Tier 2 methods are not required for AGTP compliance but SHOULD be implemented where their semantics apply.¶

1.2. Empirical Justification

Agent Transfer Protocol is motivated by measured agent performance gains. Independent benchmarking (4,800 trials across Claude, Grok, and OpenAI-family models) compared pure CRUD endpoints against a mixed catalog containing both CRUD and the semantic verbs defined in this vocabulary.¶

In mixed-paradigm conditions (Tier 1 CRUD + optional Tier 2 semantic methods), agentic verbs improved exact-match accuracy by 10–29 percentage points depending on model family (Claude +29 pp, Grok +18 pp, OpenAI +10 pp; all p<0.001). Parameter fidelity and clarification rate also improved.¶

A description-swap ablation (J1/J2) isolated the mechanism: when CRUD paths received agentic-style descriptions, performance collapsed dramatically on weaker models (Grok -39 pp, OpenAI -43 pp). When agentic paths received CRUD-style descriptions, performance held nearly flat on Claude and degraded far less on other models. This demonstrates that semantic method names encode intent more resiliently than REST-style generics, even when documentation quality varies — precisely the scenario Tier 2 methods are designed to address.¶

Two-stage discovery (as required by AGTP session establishment) incurs a modest cost (12–17 pp) but does not erase the net gain. The QUOTE method and Supported-Methods header further reduce unnecessary invocations.¶

These results validate the tiered design: Tier 1 provides the operational simplicity developers require, while the registered Tier 2 vocabulary delivers measurable agent performance and resilience. Implementations are encouraged to adopt Tier 2 methods where their domain semantics align.¶

1.3. Taxonomy

Methods are organized into six categories. Each category has a distinct operational character:¶

ACQUIRE:

Retrieve data, resources, or state without modifying it. Agents observe, locate, and extract information.¶

COMPUTE:

Process or transform information to produce a derived result. Agents analyze, classify, summarize, and generate outputs.¶

TRANSACT:

Execute operations that alter system state or complete an external commitment. Agents book, purchase, authorize, and register.¶

INTEGRATE:

Connect, synchronize, or unify data and services across system boundaries. Agents merge, link, sync, and map across silos.¶

COMMUNICATE:

Deliver signals, messages, or structured outputs to recipients. Agents notify, alert, broadcast, and reply.¶

ORCHESTRATE:

Coordinate, sequence, and manage workflows across tasks, agents, and time. Agents chain, batch, route, and schedule.¶

The category taxonomy aligns with the ACTION framework defined in [AGENTIC-API]. The INTEGRATE category is the key distinction from simpler models: cross-system unification operations are neither transactions (no external commitment) nor computations (no transformation of content into a new analytical form). They are connective operations with distinct failure modes, conflict policies, and reversibility characteristics.¶

1.4. Method Registration

All methods in this document are registered in the IANA AGTP Method Registry per the registration procedure defined in [AGTP] Section 9.2. Implementations MUST list supported Tier 2 methods in the Supported-Methods response header at session establishment. Clients SHOULD query this header before invoking Tier 2 methods.¶

3.1. FETCH

Purpose: Retrieve a specific resource by identifier. Distinguished from QUERY (which expresses an information need without specifying a location) by targeting a known resource at a known address.¶

Use case: An agent retrieves the current version of a contract document by its document ID before passing it to ANALYZE.¶

Response: Resource content with metadata. Idempotent: Yes. Error codes: 404, 403, 408.¶

3.2. SEARCH

Purpose: Execute a structured search query against a data source and return matching results. Distinguished from QUERY by accepting structured query syntax (field filters, range constraints, sort order) rather than natural language intent.¶

Use case: An agent locates all purchase orders above $50,000 in pending status, sorted by submission date.¶

Response: Result set with pagination metadata. Idempotent: Yes. Error codes: 400, 422.¶

3.3. SCAN

Purpose: Enumerate or iterate over a collection of resources, potentially applying lightweight filters. Distinguished from SEARCH by prioritizing completeness over relevance ranking.¶

Use case: An agent audits all agents in a governance zone by iterating over the full registry, page by page, to build a compliance report.¶

Response: Page of results with continuation cursor if more items remain. Idempotent: Yes. Error codes: 404, 408.¶

3.4. PULL

Purpose: Retrieve pending items from a queue or stream. Consumes items from the source; the agent takes ownership of pulled items.¶

Use case: An order-processing agent pulls the next batch of unprocessed orders from the fulfillment queue to begin packing instructions.¶

Response: List of pulled items with item IDs for acknowledgment. Idempotent: No. Error codes: 404, 409.¶

3.5. FIND

Purpose: Locate an agent, resource, or entity matching specified criteria. Distinguished from SEARCH by targeting agent or entity discovery rather than content retrieval.¶

Use case: An orchestrator locates all available analyst agents in the partner governance zone that support the VALIDATE method.¶

Response: List of matching entities with identifiers and metadata. Idempotent: Yes. Error codes: 404, 422.¶

3.6. ANALYZE

Purpose: Perform observational analysis or pattern detection on a dataset or system state. Distinguished from COMPUTE methods such as SUMMARIZE or CLASSIFY by returning findings, patterns, and anomalies rather than transforming the input. ANALYZE does not modify state.¶

Use case: An agent detects anomalous spending patterns in a quarterly expense dataset by analyzing transaction frequency and amount distributions against historical baselines.¶

Response: Findings document with detected patterns, anomaly indicators, and confidence scores per finding. Idempotent: Yes. Error codes: 400, 422, 503.¶

4.1. EXTRACT

Purpose: Pull structured information from unstructured or semi-structured source content.¶

Use case: An agent extracts party names, dates, and obligation clauses from an unstructured legal contract into a structured JSON schema.¶

Response: Extracted data conforming to the declared schema, with confidence scores per field. Idempotent: Yes. Error codes: 400, 422.¶

4.2. FILTER

Purpose: Apply criteria to a dataset and return only matching records.¶

Use case: An agent filters a product catalog to items that are in-stock, priced under $200, and tagged with a specific category.¶

Response: Filtered dataset. Idempotent: Yes. Error codes: 400, 422.¶

4.3. VALIDATE

Purpose: Check input data, a document, or a proposed action against a schema, policy, or rule set. Returns validation results but does not modify state.¶

Use case: An agent validates an invoice payload against the accounts payable schema and business rules before submitting it for payment.¶

Response: Validation result with pass/fail status, errors, and warnings. Idempotent: Yes. Error codes: 400, 422.¶

4.4. TRANSFORM

Purpose: Convert data from one format, schema, or structure to another.¶

Use case: An agent converts a vendor's proprietary order payload into the internal canonical order schema before routing to fulfillment.¶

Response: Transformed data in the target format. Idempotent: Yes. Error codes: 400, 422.¶

4.5. TRANSLATE

Purpose: Convert content between human languages.¶

Use case: An agent translates product descriptions from English into French and German to prepare a multilingual catalog update.¶

Response: Translated content with source and target language codes. Idempotent: Yes. Error codes: 400, 422.¶

4.6. NORMALIZE

Purpose: Standardize data values to a canonical form (dates, phone numbers, addresses, currency amounts, units).¶

Use case: An agent normalizes customer records ingested from three regional CRMs using different date formats, phone conventions, and address schemas into a unified canonical format.¶

Response: Normalized data with original and canonical forms. Idempotent: Yes. Error codes: 400, 422.¶

4.7. PREDICT

Purpose: Apply a model or function to input data and return a predicted output or probability estimate.¶

Use case: An agent predicts the likelihood of customer churn for each account in a segment, returning probability scores and top contributing factors for accounts above the risk threshold.¶

Response: Prediction result with confidence score and optional explanation. Idempotent: Yes. Error codes: 400, 422, 503.¶

4.8. RANK

Purpose: Sort or score a set of items by relevance, quality, or a declared criterion.¶

Use case: An agent ranks a shortlist of job candidates by predicted role fit, ordering them for recruiter review.¶

Response: Ranked list with scores per item. Idempotent: Yes. Error codes: 400, 422.¶

4.9. CLASSIFY

Purpose: Assign one or more items to categories from a known classification scheme. Distinguished from RANK (which scores and orders items) and PREDICT (which estimates future states) by assigning categorical labels to existing data.¶

Use case: An agent classifies incoming support tickets into one of eight predefined categories to route them to the appropriate team.¶

Response: Classification result with assigned labels and confidence scores per item. Idempotent: Yes. Error codes: 400, 422.¶

4.10. CALCULATE

Purpose: Perform numeric, logical, or financial computations on structured inputs. Distinguished from PREDICT (probabilistic) and EVALUATE (qualitative comparison) by operating on deterministic mathematical or rule-based logic.¶

Use case: An agent calculates the total cost of a procurement order including applicable taxes, shipping fees, and bulk discount rules.¶

Response: Computation result with inputs echoed, formula applied, and output value. Idempotent: Yes. Error codes: 400, 422.¶

4.11. EVALUATE

Purpose: Compare a target against benchmarks, rules, standards, or criteria and return a qualitative or scored assessment. Distinguished from VALIDATE (pass/fail against a schema) by returning a graded assessment with explanatory context.¶

Use case: An agent evaluates a vendor proposal against a weighted scorecard covering price, delivery timeline, compliance posture, and references, returning a composite score and narrative rationale.¶

Response: Evaluation result with per-criterion scores, composite score, and narrative rationale. Idempotent: Yes. Error codes: 400, 422.¶

4.12. GENERATE

Purpose: Produce textual, structured, or code output from a source specification, dataset, or prompt. Distinguished from SUMMARIZE (which condenses existing content) by creating new content that did not exist in the source input.¶

Use case: An agent generates API documentation from an OpenAPI specification, producing endpoint descriptions, parameter tables, and example payloads.¶

Security note: GENERATE with output_type: code MUST require human or governance review before passing to RUN. Implementations providing a pipeline from GENERATE to RUN MUST insert an explicit VALIDATE or approval step between them.¶

Response: Generated content with output_type field. Idempotent: Yes (for identical inputs). Error codes: 400, 422, 503.¶

4.13. RECOMMEND

Purpose: Generate a ranked list of suggestions, options, or actions tailored to a declared context, goal, and constraints. Distinguished from RANK (which scores a provided list) by generating the candidate set itself based on the agent's understanding of the context.¶

Use case: An agent recommends three product accessories to a customer based on their purchase history, stated budget, and current browsing context.¶

Response: Ordered list of recommendations with scores, rationale, and confidence per item. Idempotent: Yes. Error codes: 400, 422, 503.¶

5.1. QUOTE

Purpose: Pre-flight cost estimation for a proposed method invocation. The requesting agent submits a proposed method call; the server returns a cost estimate without executing the method. Servers supporting Budget-Limit SHOULD implement QUOTE.¶

Use case: An agent estimates the token cost of a large-document SUMMARIZE operation before committing, to verify it stays within the session budget.¶

Response: Cost estimate in Cost-Estimate response header and response body, with confidence range. Idempotent: Yes. Error codes: 400, 404, 422.¶

AGTP/1.0 200 OK
Task-ID: task-quote-01
Cost-Estimate: tokens=8500 compute-seconds=2.3
Content-Type: application/agtp+json

{
  "status": 200,
  "task_id": "task-quote-01",
  "result": {
    "method": "QUERY",
    "estimated_cost": {
      "tokens": {"min": 7200, "expected": 8500, "max": 11000},
      "compute_seconds": {"min": 1.8, "expected": 2.3, "max": 4.0}
    },
    "confidence": 0.82
  }
}

5.2. REGISTER

Purpose: Create a new record, entity, or subscription in a target system.¶

Use case: An onboarding agent registers a new employee in the identity management system, the access control system, and the payroll platform as three sequential REGISTER calls within a single session.¶

Response: Registration receipt with new entity identifier. Idempotent: No (use idempotency_key for safe retry). Error codes: 409, 422.¶

5.3. SUBMIT

Purpose: Deliver data, a document, or a work item to a processing system or queue for handling.¶

Use case: An agent submits a completed expense report to the finance system's approval queue for human review.¶

Response: Submission receipt with tracking identifier. Idempotent: No. Error codes: 400, 409, 503.¶

5.4. AUTHORIZE

Purpose: Approve permissions, credentials, or access for an entity, action, or resource. Distinguished from VALIDATE (which checks conformance) by issuing an active grant rather than a check result.¶

Use case: An identity agent authorizes a contractor's access request for a restricted project repository, generating a scoped access token with a defined expiry.¶

Security note: AUTHORIZE MUST carry authorization:grant in the agent's Authority-Scope. The granted scope MUST NOT exceed the authorizing agent's own declared scope (the anti-laundering constraint from [AGTP] Section 7.2.7 applies to all authorization grants). 451 Scope Violation MUST be returned if either constraint is violated. Idempotent: No. Error codes: 403, 409, 451.¶

5.5. CANCEL

Purpose: Revoke or reverse a previously scheduled or committed transaction. Distinguished from lifecycle operations (agent suspension or revocation) by targeting a specific transaction, booking, or reservation rather than an agent's registry state.¶

Use case: An agent cancels a conference room booking on behalf of the principal, releasing availability and notifying other attendees.¶

Response: Cancellation receipt with confirmation and any refund details. Idempotent: No (cancellation of an already-cancelled item SHOULD return 409 Conflict). Error codes: 404, 409, 422.¶

5.6. TRANSFER

Purpose: Move ownership or custody of a resource from one principal or agent to another.¶

Use case: An agent transfers custodianship of a finalized contract to the legal archive system, removing it from the active working directory and updating ownership records.¶

Security note: TRANSFER MUST verify that the requesting agent holds Authority-Scope for the resource type being transferred. 451 Scope Violation MUST be returned if not. Idempotent: No. Error codes: 403, 404, 409, 451.¶

5.7. PURCHASE

Purpose: Execute a financial transaction to acquire a resource, service, or allocation.¶

Use case: A travel agent purchases a confirmed airline seat after the principal approves the itinerary, carrying the principal's payment method identifier.¶

Security note: PURCHASE MUST carry payments:purchase in the agent's Authority-Scope. 451 Scope Violation MUST be returned if absent. PURCHASE MUST validate against the Budget-Limit header if present. Idempotent: No. Error codes: 403, 409, 451, 452.¶

5.8. SIGN

Purpose: Apply a cryptographic signature to a document, assertion, or data artifact using the agent's or principal's signing key.¶

Use case: An agent signs a finalized service agreement on behalf of the principal using the organization's governance key, producing a timestamped, non-repudiable signature record.¶

Response: Signed artifact with signature metadata. Idempotent: No (signatures include timestamps). Error codes: 403, 422.¶

5.9. LOG

Purpose: Write a structured record to an audit trail or operational log. Agents SHOULD use LOG for governance-significant events rather than embedding log data in other method payloads.¶

Use case: A financial agent logs each step of a multi-stage wire transfer approval workflow as discrete LOG events, creating an audit trail for compliance reporting.¶

Response: Log receipt with record identifier. Idempotent: No. Error codes: 400, 503.¶

5.10. PUBLISH

Purpose: Make content or data available to a defined audience or channel.¶

Use case: An agent publishes a finalized research summary to the organizational knowledge base, making it discoverable to all users with knowledge:read scope.¶

Response: Publication receipt with content identifier and channel confirmation. Idempotent: No. Error codes: 400, 403, 409.¶

6.1. MERGE

Purpose: Combine two or more datasets, documents, or resources into a unified result, resolving conflicts according to a declared policy.¶

Use case: An agent merges duplicate customer profiles detected across two acquired subsidiary databases into a single canonical record, applying a "most recently updated" conflict resolution strategy.¶

Response: Merged result with conflict report if applicable. Idempotent: No. Error codes: 400, 409, 422.¶

6.2. LINK

Purpose: Create a persistent association or relationship between two entities in a system of record. Distinguished from MAP (which defines structural schema correspondence) by creating a relationship record between specific instances.¶

Use case: An agent links a user's enterprise SSO identity to their account in a third-party analytics platform, establishing a persistent cross-system identity association.¶

Response: Link record with relationship identifier. Idempotent: No. Error codes: 404, 409.¶

6.3. SYNC

Purpose: Reconcile the state of a local resource with a remote authoritative source, aligning records bidirectionally or unidirectionally.¶

Use case: An agent synchronizes the local product inventory cache with the warehouse management system, pulling delta changes since the last sync timestamp and pushing any local updates that occurred offline.¶

Response: Sync receipt with change summary. Idempotent: No. Error codes: 404, 409, 503.¶

6.4. IMPORT

Purpose: Bring external data into the agent's operational context or a designated storage target from a non-AGTP-native source. Distinguished from FETCH (which retrieves a known resource by ID from an AGTP-native address) by ingesting from external systems.¶

Use case: An agent imports a CSV export of contact records from a legacy CRM into the new system, resolving format differences and logging import statistics.¶

Response: Import receipt with record count and any conflict resolutions. Idempotent: No. Error codes: 400, 409, 422.¶

6.5. MAP

Purpose: Define a structural correspondence between two schemas, data models, or field sets. Distinguished from TRANSFORM (which applies a mapping to convert a specific payload) by producing a reusable mapping definition rather than executing a one-time conversion.¶

Use case: An agent generates a field-level mapping between an external vendor's order schema and the internal canonical order schema, producing a reusable mapping document that subsequent TRANSFORM calls can reference.¶

Response: Mapping document with per-field correspondence and confidence scores. Idempotent: Yes. Error codes: 400, 422.¶

6.6. CONNECT

Purpose: Establish a conduit, session, or integration channel between two systems, services, or agents. Distinguished from LINK (which creates a record-level association) by establishing an active communication or data pathway.¶

Use case: An agent establishes a streaming data connection between a sensor telemetry feed and the observability platform, configuring the channel parameters and returning a handle for subsequent operations.¶

Response: Connection handle with connection ID and channel metadata. Idempotent: No. Error codes: 404, 409, 503.¶

6.7. EMBED

Purpose: Insert one component, dataset, or agent capability into another as a nested or composed element. Used when an integration requires one system to host or expose the functionality of another within its own operational context.¶

Use case: An agent embeds a third-party risk-scoring model's output into the organization's credit decisioning workflow as a named sub-component, making its scores available inline within decisioning records.¶

Response: Embedding receipt with binding reference and version. Idempotent: No. Error codes: 400, 404, 409.¶

7.1. ALERT

Purpose: Send a high-priority, time-sensitive message to a recipient or group. Distinguished from NOTIFY (Tier 1, general-purpose async push) by signaling urgency and requiring acknowledgment handling if not acknowledged within a declared timeout.¶

Use case: An infrastructure monitoring agent sends a critical alert to the on-call team when database latency exceeds the SLA threshold, carrying metric value, threshold, and recommended escalation path.¶

Response: Alert delivery receipt with alert ID. Implementations SHOULD track acknowledgment state and trigger escalation on timeout. Idempotent: No. Error codes: 400, 404, 503.¶

7.2. BROADCAST

Purpose: Disseminate information simultaneously to multiple recipients or a defined group. Distinguished from NOTIFY (targeted at a specific recipient) by addressing an audience rather than an individual and not expecting individual responses.¶

Use case: An agent broadcasts a scheduled maintenance window notification to all registered subscriber endpoints across three governance zones.¶

Response: Broadcast receipt with message ID and delivery count. Idempotent: No. Error codes: 400, 403, 503.¶

7.3. REPLY

Purpose: Provide a direct, traceable response to an incoming NOTIFY, ESCALATE, or QUERY that explicitly requested a reply. Creates a response thread linked to the originating task.¶

Use case: A human-in-the-loop handler sends a REPLY to a pending ESCALATE, providing a decision and carrying the original escalation ID for thread continuity.¶

Response: Reply receipt. Idempotent: No. Error codes: 404, 400.¶

7.4. SEND

Purpose: Deliver a message or payload to a recipient through a declared external channel. Distinguished from NOTIFY (AGTP-native async push) by targeting non-AGTP delivery channels.¶

Use case: An agent sends an email confirmation to a customer's registered email address after completing a booking, using a declared template identifier.¶

Response: Delivery receipt with channel-specific confirmation. Idempotent: No. Error codes: 400, 404, 503.¶

7.5. REPORT

Purpose: Generate and deliver a structured summary or analysis document to a principal or system. Distinguished from NOTIFY (raw payload push) by producing a formatted output artifact with declared structure.¶

Use case: An agent generates and delivers a weekly sales performance summary to the sales leadership distribution list in PDF format.¶

Response: Report delivery receipt with document identifier. Idempotent: No. Error codes: 400, 422, 503.¶

8.1. CHAIN

Purpose: Link a defined sequence of AGTP method invocations into a composite workflow where each step's output may be used as input for subsequent steps. Distinguished from BATCH (which executes unrelated tasks in parallel) by enforcing sequential dependency between steps.¶

Use case: An agent chains SEARCH, followed by ANALYZE, followed by REPORT into a single declared workflow for weekly competitor intelligence, where each step uses the prior step's output.¶

Response: Execution receipt with chain ID and per-step status. Idempotent: No. Error codes: 400, 408, 422.¶

8.2. BATCH

Purpose: Group multiple independent AGTP method calls for concurrent or bulk execution, reducing round-trip overhead. Distinguished from CHAIN (sequentially dependent) by executing steps independently without output dependency.¶

Use case: An agent batches twenty TRANSLATE calls for product descriptions into a single BATCH invocation, reducing latency from twenty serial round-trips to a single parallel execution.¶

Response: Batch result containing individual outcome per call. Idempotent: Conditional (depends on idempotency of component calls). Error codes: 400, 408, 422.¶

8.3. MONITOR

Purpose: Establish ongoing observation of an agent, resource, or condition. Returns a subscription handle; the AGTP server delivers updates via NOTIFY when observed conditions change.¶

Use case: An agent establishes a monitor on the behavioral trust score of a set of sub-agents, receiving NOTIFY updates when any agent's score drops below the declared threshold.¶

Response: Monitor subscription receipt with subscription identifier. To cancel monitoring, send NOTIFY with event type monitor_cancel and the subscription identifier. Idempotent: No. Error codes: 404, 400.¶

8.4. ROUTE

Purpose: Direct a task, message, or work item to the appropriate handler or agent based on declared routing criteria.¶

Use case: An agent routes an incoming customer inquiry to the appropriate specialized handling agent based on topic classification, account tier, and declared availability of candidate handlers.¶

Response: Routing decision with selected handler identifier. Idempotent: No. Error codes: 404, 422, 503.¶

8.5. RETRY

Purpose: Re-attempt a previously failed AGTP method invocation using a prior Task-ID. Distinguished from a new invocation by carrying the original Task-ID for idempotency checking at the server.¶

Use case: An agent retries a failed BOOK call after a 503 Unavailable, carrying the original Task-ID so the booking system can detect and suppress any duplicate if the original call was partially processed.¶

Response: New task receipt linked to the original task. Idempotent: Yes (for idempotent original methods); No for non-idempotent originals. Error codes: 404, 409.¶

8.6. PAUSE

Purpose: Temporarily halt a scheduled or repeating workflow without terminating it. Distinguished from SUSPEND (Tier 1, which targets a session) by targeting a workflow or schedule record.¶

Use case: An agent pauses a recurring data pipeline workflow during a maintenance window, setting a resume timestamp at the scheduled end of the window.¶

Response: Pause receipt with workflow status. Idempotent: No. Error codes: 404, 409.¶

8.7. RESUME

Purpose: Restart a paused workflow or a suspended session.¶

Use case: An agent resumes a paused procurement workflow after the required human approval has been received, using the resumption nonce issued at PAUSE time.¶

Response: Resumption receipt with current workflow status. Idempotent: No. Error codes: 404, 408, 409.¶

8.8. RUN

Purpose: Execute a named, registered procedure or automation script. Implementations MUST NOT accept free-form execution strings; the procedure MUST be identified by a registered procedure_id.¶

Use case: An agent runs a registered data quality validation procedure against a freshly imported dataset, passing the dataset ID as input and receiving a structured pass/fail report.¶

Security note: RUN MUST require activation:run in the agent's Authority-Scope. Free-form execution strings MUST NOT be accepted under any circumstances. Idempotent: No. Error codes: 403, 404, 408, 422, 451.¶

8.9. CHECK

Purpose: Query the status or health of an agent, resource, workflow, or external dependency. Returns a structured status response.¶

Use case: An agent checks the availability of all three downstream payment processors before initiating a purchase, using depth: shallow to minimize latency.¶

Response: Status document with health indicators per component. Idempotent: Yes. Error codes: 404, 408.¶

10.1. PURCHASE Authorization

The PURCHASE method carries financial consequences and MUST be subject to strict scope enforcement. Implementations MUST reject PURCHASE requests that do not carry payments:purchase in the Authority-Scope header with 451 Scope Violation. Budget-Limit validation MUST occur before execution; 452 Budget Exceeded MUST be returned if the purchase amount would exceed the declared budget.¶

10.2. AUTHORIZE Scope Constraint

The AUTHORIZE method grants permissions to other entities and is subject to the same anti-laundering constraint that applies to DELEGATE in the Tier 1 core: the scope granted by an AUTHORIZE call MUST NOT exceed the authorizing agent's own Authority-Scope. Any implementation receiving an AUTHORIZE request where the grant scope exceeds the authorizing agent's declared scope MUST return 451 Scope Violation and MUST log the event.¶

10.3. RUN Method Safety

The RUN method is the highest-risk Tier 2 method. Implementations MUST maintain a registry of permitted procedure_id values and MUST NOT execute procedures not in that registry. Free-form execution strings MUST be rejected. Each RUN invocation MUST be logged with the full procedure_id, input parameters, and executing Agent-ID.¶

10.4. GENERATE Code Output Safety

When GENERATE is invoked with output_type: code, the resulting code MUST NOT be passed directly to RUN without human or governance review. Implementations providing a pipeline from GENERATE to RUN MUST insert an explicit VALIDATE or approval step between them.¶

10.5. MONITOR Callback Verification

The callback parameter in MONITOR specifies an AGTP endpoint to receive updates. Implementations MUST verify that the callback endpoint is reachable and that the requesting agent has authority to receive notifications at that endpoint before establishing a monitoring subscription. Unverified callbacks are a potential exfiltration vector.¶

10.6. TRANSFER and Ownership Chain Integrity

TRANSFER operations modify resource ownership and create audit obligations. Implementations MUST record the complete ownership chain for each transferred resource in the governance audit trail. The Attribution-Record for TRANSFER MUST include both from_principal and to_principal.¶

12.1. Normative References

[AGTP]

Hood, C., "Agent Transfer Protocol (AGTP)", Work in Progress, Internet-Draft, draft-hood-independent-agtp-02, 2026, <https://datatracker.ietf.org/doc/html/draft-hood-independent-agtp-02>.

[RFC2119]

Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, DOI 10.17487/RFC2119, March 1997, <https://www.rfc-editor.org/rfc/rfc2119>.

[RFC8126]

Cotton, M., Leiba, B., and T. Narten, "Guidelines for Writing an IANA Considerations Section in RFCs", BCP 26, RFC 8126, DOI 10.17487/RFC8126, June 2017, <https://www.rfc-editor.org/rfc/rfc8126>.

[RFC8174]

Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174, May 2017, <https://www.rfc-editor.org/rfc/rfc8174>.

12.2. Informative References

[AGENTIC-API]

Hood, C., "AgenticAPI: A Task-Centric Framework for Scalable Agent Integrations", 2025, <https://agenticapi.io>.

[RFC6749]

Hardt, D., Ed., "The OAuth 2.0 Authorization Framework", RFC 6749, DOI 10.17487/RFC6749, October 2012, <https://www.rfc-editor.org/rfc/rfc6749>.