ADR#0070: An Agent Declares What It Offers and Whose Authority It Carries; Throughput Ceilings Stay on the Policy Plane
Context
ADR#0025 admits a datum into AgentConfiguration only when it is the Agent's own behavior or dependency declaration, every session pinned to the revision should begin with it, and changing it should invalidate behavioral verification. Everything the configuration holds today looks inward: runtime settings, skill pins, and tool, delegate, memory, and connection declarations under ADR#0062 and ADR#0063 all describe what the Agent needs. Nothing describes what it offers.
That gap blocks a decision already made. ADR#0012 allows an A2A agent card to be exported as a catalog entry with media type application/a2a-agent-card+json. The A2A specification requires a card to list the skills the agent offers, its default input and output modes, and its capabilities, including streaming, push notifications, and state transition history. The card's name and description have sources in the Agent and its configuration. Its skills, modes, and capabilities have no source anywhere in the schema, so an export would have to invent them.
A second inward gap concerns whose authority a run carries. Some Agents act only as themselves. Others exist to act for an authenticated human, and a session that starts without one is a different run, not the same run with less context. Nothing in the configuration says which kind of Agent it is, so admission cannot reject a session that starts in the wrong mode.
The policy plane has its own gap. policy/agent_budgets caps money per Agent over a calendar month or a lifetime. An agentic loop can make unbounded model calls inside one session, and the platform has no ceiling on a single session's spend and no throughput control at all: no token or request ceiling over a window and no cap on how many model calls one session may make. ADR#0025 Decision 5 already lists "budget, token, turn, concurrency, and delegation ceilings" among the data that never enters an AgentConfiguration, and ADR#0062 gives budgets their own control plane. The question is only where the new ceilings live on that plane.
The agent platform research index records the products studied for the definition, session, and policy split this record extends.
Decision
1. AgentConfiguration.interface declares what the Agent offers
AgentConfiguration gains AgentInterface interface. An AgentInterface carries:
offered_skills, each anOfferedSkillwith a stableskill_key, adisplay_namethat nothing resolves per ADR#0040, adescription,tags,examples, and optional per-skill input and output media type overrides that otherwise inherit the interface defaults;default_input_media_typesanddefault_output_media_types; andInteractionCapabilities, withstreaming,push_notifications, andstate_transition_history.
The interface passes every statement of the ADR#0025 ownership test. It is the Agent's own declaration of behavior. Every session pinned to the revision begins with the same offer, and session admission validates inbound and outbound parts against it. Changing it changes what callers may ask for and what they get back, which invalidates behavioral verification and requires a new revision.
It is the protocol-neutral source from which the catalog layer of ADR#0012 derives an A2A card. The interface is not A2A-shaped on purpose, consistent with that record's rule that the neutral model must not reuse A2A identifiers as its domain.
An OfferedSkill is the opposite direction from SkillDeclaration. A SkillDeclaration pins content the Agent consumes; an OfferedSkill advertises an ability callers consume. Advertising grants no authority and resolves no dependency.
An absent interface means the Agent offers no external interface and cannot be exported as a callable catalog entry.
The interface deliberately omits the remaining card fields, because each already has an owner:
- endpoints and transports are platform bindings under ADR#0016;
- security schemes belong to ADR#0017;
- card signatures are produced at export with the Agent's own key under ADR#0036, so no stored signature can drift from the card it signs; and
- the provider organization derives from the hierarchy
parent, which ADR#0025 assigns to the hierarchy plane.
2. AgentConfiguration.execution_subject declares whose authority a session carries
AgentConfiguration gains ExecutionSubject execution_subject, with the values AGENT_ONLY, DELEGATED_PRINCIPAL_ONLY, and AGENT_OR_DELEGATED_PRINCIPAL. Absent means AGENT_ONLY.
Whether a run carries a human principal changes what the Agent does and what verification must cover, so the declaration is configuration-owned. Session admission enforces it: a session that does not match the declared subject is rejected, and under AGENT_OR_DELEGATED_PRINCIPAL the session records which mode applied.
The declaration grants no authority. It names no principal and does not say who may invoke the Agent, which ADR#0025 keeps on authentication and policy. The identity binding itself stays on the identity plane under ADR#0036 and ADR#0037. The configuration says only which binding a session must present.
3. AgentBudgetPeriod.SESSION caps money per Session
policy/agent_budgets gains AGENT_BUDGET_PERIOD_SESSION beside the calendar month and lifetime periods. It caps all spending attributable to one Session from its start through its terminal state. Spending a Session delegates to descendant Sessions counts against its root. Each Session's counter starts at zero and is never shared with another Session.
A per-session money ceiling is policy, not configuration: it is a funder's stance, it changes at the funder's cadence, and changing it must not mint a revision. It therefore joins the existing budget policy rather than the revision. Admission evaluates it in the same conjunction as the other periods, and it neither resets nor replaces their counters.
4. policy/agent_rate_limits adds throughput ceilings
A new policy module policy/agent_rate_limits holds an AgentRateLimitPolicy per Agent. The policy carries:
AgentRateLimitentries, each keyed by a scope (AGENTorSESSION) and a meter (TOKENSorMODEL_REQUESTS), with amaximumand a fixedwindow; and- an
AgentSessionCallLimit, a cap on total model requests per Session with no window.
The module follows the shape of policy/agent_budgets: initialize, set, and clear commands with matching events, and the same role as a provisioning participant. Counters live in authoritative accounting history and never in the policy stream, so setting, clearing, or replaying a policy cannot reset them.
The platform had money ceilings only. Throughput is a separate axis, because a cheap model can still exhaust an upstream quota or run a loop indefinitely, and ADR#0025 and ADR#0062 already place ceilings on the policy plane. A separate module keeps money and throughput independently settable, since they are often owned by different people.
5. AgentProfile.documentation_url is presentation data
The catalog profile gains documentation_url, which A2A cards also carry. It accepts only credential-free absolute HTTPS URLs, under the same admission rule as Avatar.https_url.
It belongs on the profile, not the revision. A documentation link changes nothing any session does, so it fails the ADR#0025 verification test, and updating it must not require a proposal.
6. Rejected or deferred
- Guardrail bindings. No guardrail resource exists for a binding to reference. A binding without a target is an unverifiable string. ADR#0071 records the trigger and the shape such a binding will take.
- A paused but not archived lifecycle state. Pausing changes the registry lifecycle, which is a separate lifecycle decision rather than an interface or limit decision. ADR#0071 meets the need with a hold on the policy plane instead of a lifecycle state.
- An operational kill-switch webhook. Stopping a running Agent belongs to a runtime or session plane, and an outbound webhook only matters when the platform proxies an externally hosted Agent.
- Per-agent trace-enforcement toggles. Trace context is already mandatory under ADR#0042. A toggle could only weaken a platform guarantee.
- Pricing fields. What a caller pays belongs to the settlement plane, not to the Agent's offer or to its spending ceilings.
Consequences
- An Agent without an interface is not exportable as a callable catalog entry. Export is a consequence of a declaration, never a default.
- Offered skill keys are unique within an interface and stable across revisions, so a caller or catalog can follow one skill through the Agent's history.
- A session-scoped budget or rate limit never shares counters between Sessions, and a root Session's budget counter includes its descendants.
- Each scope and meter pair appears at most once in an
AgentRateLimitPolicy. - An absent budget or rate limit policy removes only that Agent-specific ceiling. It is never a zero cap and never bypasses another scope's limit or an authorization check.
- Configuration digests minted before this change remain valid, since the new fields are additive at fresh tags and absent values serialize to zero bytes.
- The catalog export layer of ADR#0012 must map
AgentInterface, the Agent's name and description, andAgentProfile.documentation_urlto the A2A card shape, and must produce the card's endpoints, security, signatures, and provider from their own planes. - Session admission must enforce
execution_subjectand must validate parts against the interface's media types, including per-skill overrides. Until it does, both fields are reviewable declarations without enforcement. - Authoritative accounting must attribute spend, tokens, and model requests by Session, including descendant Sessions under their root, before the session period or the session scope can admit or reject anything.
References
- ADR#0012: ARD-Compatible Discovery Catalog
- ADR#0016: Protocol Buffers RPC over NATS micro Binding
- ADR#0017: AAuth Agent Authentication over a Trogon NATS PoP Binding
- ADR#0025: Agent Definition Data Ownership
- ADR#0036: Agent Self-Certifying Cryptographic Identity
- ADR#0037: Agent Identity Governance
- ADR#0042: NATS Trace Context and Message Path Tracing
- ADR#0062: Runtime-Owned Settings and Platform Declarations
- ADR#0063: An Agent Declares the Connections It Needs
- Agent2Agent (A2A) specification
- Agent platform research index
