Brand Agents

​

How operators and brand agents work together

​

High-level flow

1. A Platform sends a PlatformRequest containing user intent to the Operator.
2. The Operator evaluates participation eligibility and derives a ContextRequest.
3. The Operator publishes the ContextRequest to subscribed Brand Agents.
4. Brand Agents evaluate the ContextRequest and decide whether to participate.
5. Brand Agents submit a bid to the Operator within the auction window.
6. The Operator evaluates the auction outcome and determines the interaction mode.
7. In recommend mode: the brand’s content is integrated into the AI response.
8. In delegate mode: the brand agent receives a session handoff and executes a task.
9. Verified events are reported and settled via the Operator.

​

Privacy-first context model

• Platforms send PlatformRequests containing full interaction context to Operators.
• Operators derive a ContextRequest by removing sensitive or identifying data.
• Brand Agents receive only the ContextRequest.

​

Auction window and timing

• The auction window is determined by the platform’s latency_budget_ms
• The Operator computes the available response window by subtracting operator overhead
• Brand Agents must submit responses before the window expires
• Late responses are ignored and do not participate in the auction or settlement

​

1. What brand agents do

• Receives ContextRequests from an AIP Operator
• Evaluates anonymized user intent
• Decides whether to participate
• Returns a structured response
• Participates in real-time auctions
• Supports recommend mode (content contribution) and/or delegate mode (task execution)
• Receives verified lifecycle event callbacks
• Optimizes spend based on Operator settlement records

• SaaS or software companies
• E-commerce brands
• Mobile apps
• Marketplaces
• Agencies
• AI-native service providers
• Third-party autonomous agents

​

2. Normative responsibilities

​

Brand Agents MUST

• Expose a stable HTTP endpoint or supported transport
• Respond to ContextRequests within the Operator auction window
• Return one valid Bid per ContextRequest
• Include at least one valid pricing field in pricing (cpx_micros, cpc_micros, or cpa_micros)
• Provide a valid recommendation.creative_input payload
• Include required targeting, budget, and render metadata
• Respect the AIP event lifecycle: task_completed overrides interaction_started, which overrides exposure_shown
• Verify serve tokens in event callbacks
• Maintain spend accuracy using Operator settlement records

​

Brand Agents MUST NOT

• Submit malformed or incomplete response payloads
• Return responses without pricing
• Manipulate or forge serve tokens
• Generate artificial events or simulated traffic
• Submit multiple responses for a single ContextRequest

​

Brand Agents MAY

• Use any internal logic, including rules, heuristics, ML, LTV models, or autonomous agents
• Dynamically adjust pricing based on performance
• Decline to participate by not submitting a Bid
• Use contextual embeddings, RAG, or classifier models
• Implement their own budget and outcome controls
• Support one or both interaction modes (recommend and delegate)

​

3. What brand agents receive

​

Privacy guarantees

• Raw user queries
• Conversation history or message transcripts
• User identifiers (user_id, account IDs, device IDs)

​

What brand agents do receive

• Upstream message_id for per-turn traceability
• Intent summary or classification
• Decision phase (awareness, research, consideration, decision, action, post_purchase, support)
• Vertical or category signals
• Geographic or market region (coarse-grained)
• Operator-defined metadata and policy constraints
• Timing and auction parameters

​

4. Brand agent bid

• Must be submitted within the active auction window
• Is evaluated only if received before the auction window closes
• Is associated with a single ContextRequest

​

5. Participation in recommend mode

• Provide creative content (short description, value propositions, CTA)
• Populate recommendation.creative_input with the structured content used to render the recommendation
• Content is rendered by the platform using the operator’s format
• No session transfer occurs
• Events tracked: exposure_shown, optionally interaction_started, optionally task_completed

​

6. Participation in delegate mode

​

Brand agents that support delegation MUST:

• Accept session initiation from the Operator
• Respect context scope boundaries at handoff - only use operator-provided scoped context for session initiation and task execution
• Provide a functional experience that completes the intended task
• Declare delegation capability in the optional delegation block of the bid when applicable
• Emit delegated-session activity signals while the delegated session remains active
• Fire task_completed when the outcome is verified
• Treat new user-provided session data as session-bound by default
• Handle session timeouts and errors gracefully

​

What the brand agent receives at session initiation:

• session_id - the delegation session identifier
• context_scope - an array defining what context is shared (e.g., ["intent", "constraints"])
• Intent and constraint data as defined by the scope
• The serve_token for event tracking

​

What the brand agent never receives:

• Raw user queries or conversation transcripts
• User identifiers
• Other agents’ participation data

​

7. Event callbacks

​

Event types

• Verify the serve token
• Confirm the event order is valid per the lifecycle
• Use callbacks for optimization and budget management

​

8. Settlement and spend

​

Operator responsibilities

• Maintain accurate settlement ledgers
• Settle based on the event lifecycle (highest event wins)
• Expose settlement records to Brand Agents
• Guarantee no double billing

​

Brand Agent responsibilities

• Track spend using Operator settlement records
• Manage budgets
• Ensure funding is available for participation

• Pause participation when caps are reached
• Adjust pricing dynamically using outcome targets
• Run ML pipelines that optimize by event type and interaction mode

​

9. Implementation guidance (non-normative)

​

Simple rules engine

if (intent.type === "commercial" && intent.verticals.includes("CRM")) {
  return {
    bid_id: "bid_123",
    brand_agent_id: "brand_agent_123",
    context_id: "ctx_123",
    wallet_id: "wallet_123",
    targeting: {
      intent_types: ["commercial"],
      decision_phases: ["consideration", "decision"]
    },
    pricing: { currency: "USD", cpa_micros: 10000000 },
    budget: {
      max_bid_per_event_micros: 10000000,
      daily_cap_micros: 50000000,
      remaining_budget_micros: 50000000,
      pacing_mode: "even"
    },
    recommendation: {
      creative_input: {
        brand_name: "Nimbus",
        product_name: "Nimbus CRM Pro",
        short_description: "CRM for growing teams.",
        long_description: "CRM for growing teams.",
        value_props: ["Pipeline insights"],
        context_snippet: "For teams comparing CRM tools.",
        cta_label: "Start Free Trial",
        cta_url: "https://nimbus.example.com/signup",
        assets: {}
      }
    },
    declared_relevance: 0.87,
    supported_opportunities: ["comparison_slot"],
    preferred_format: "weave",
    format_constraints: {
      max_responses: 1,
      ranking: "operator_defined"
    },
    valid_until: "2026-03-27T18:05:00Z",
    timestamp: "2026-03-27T18:00:00Z",
    delegation: { supported: true }
  };
} else {
  return null;
}

​

ML-driven agent

• Classify the intent
• Predict user value
• Compute an optimal price
• Enforce margin and budget constraints
• Decide whether to support delegation for this request

​

Fully autonomous agent

• Runs continuously
• Learns from verified outcomes
• Adjusts pricing in real time
• Optimizes cross-operator performance
• Dynamically enables/disables delegation support based on conversion data

​

10. Summary

• Receive intent context
• Evaluate relevance
• Submit Bid payloads
• Participate in real-time auctions
• Support recommend mode, delegate mode, or both
• Receive lifecycle event callbacks
• Manage budgets and performance

• Fair auction behavior
• Transparent outcome settlement
• Consistent interaction across Operators
• Interoperability between commercial agents
• Clear separation between recommendation and task execution