RFC: Intelligence agent type for discoverable market intelligence and strategic decision support

[skreepatch]

Summary

Propose a new intelligence agent type for AdCP that makes market intelligence a discoverable capability within the protocol — rather than something buyer agents must obtain through proprietary integrations outside the standard.

Motivation

AdCP's core value is standardizing how agents discover and interact with capabilities. Today that discovery model covers products (sales), execution (buying), contextual data (signals), creative production (creative), and compliance (governance).

As autonomous buyer agents become more capable, they will also need to discover and consume market intelligence in a standardized way. Today there is no protocol-level mechanism for this — buyers who want strategic context must build bilateral integrations outside AdCP, which is precisely the fragmentation the protocol exists to solve.

The missing layer is interpretation. Signals describe reality. Intelligence interprets reality. Buyer agents use that interpretation to make better decisions.

This RFC proposes formalizing that layer as a first-class agent type.

The architectural gap

A buyer agent today can:

• Discover available inventory (get_products)
• Activate contextual signals (get_signals, activate_signal)
• Execute a media buy (create_media_buy)

What it cannot do is discover, in a standardized way, an agent that can tell it: what is the competitive pressure in this market slice right now? What is the market trajectory? What is my optimal position given current conditions?

Without a standard type, intelligence providers remain outside the protocol — proprietary, non-discoverable, and non-interoperable.

The core architectural pattern

The most important design property of a market intelligence agent is the shared model / individualized perspective pattern:

• Multiple participants contribute observations to a common abstract market model
• Each participant receives only an interpretation relevant to their own position and interests
• No participant's competitive data is exposed to others

This is a general architectural pattern, not an implementation detail. It is what makes market intelligence safe to standardize: the shared model enables useful aggregate intelligence; individual scoping preserves competitive isolation.

Proposed agent type

A new top-level agent type: intelligence

An intelligence agent provides advisory operations that interpret market conditions for buyer agents. It does not execute transactions — it informs the decisions that lead to them.

Rather than prescribing a fixed task interface upfront, this RFC proposes that intelligence agents may expose advisory operations — such as get_market_context, get_market_guidance, or evaluate_position — and that the working group evolve the exact interface through implementation experience.

Discovery would follow the existing pattern: intelligence agents declare their capabilities in get_adcp_capabilities, including the market slices they cover, allowing buyer agents to find relevant providers through standard protocol discovery.

Relationship to existing types

• Not signals — signals are observations about reality; intelligence is interpretation of reality for strategic use
• Not buying — intelligence informs decisions but does not execute them
• Not governance — governance enforces constraints; intelligence provides strategic guidance within those constraints
• Complements all three — an intelligence agent may consume signals, inform buying decisions, and operate within governance boundaries

Open questions for working group

1. Should intelligence be a new top-level agent type, or a specialism within an existing type (e.g., a signals agent with an advisory specialism)?
2. How should competitive isolation be enforced — at the protocol level, or left to the intelligence agent's implementation?
3. Should market slices be standardized within the AdCP taxonomy, or remain provider-defined?
4. How should buyer agents signal their own context to an intelligence agent without exposing competitive data to other participants?
5. Should strategic guidance be strictly advisory, or should it optionally produce structured parameters a buyer agent can pass to create_media_buy?
6. What data minimization requirements apply to aggregate market model construction?
7. Is the shared model / individualized perspective pattern a general primitive worth formalizing beyond this agent type?

Related

• Discussion: [Progressive Agentic Adoption — Architecture and Technical Standards WG]
• Discussion questions fix: remove empty line before frontmatter in creative-formats.md #5 and Rename Audience Activation Protocol to Signals Activation Protocol #6 in the related working group post anticipate this architectural boundary

[bokelley]

Triage

Classification: Feature request — RFC
Bucket(s): spec / protocol
Status: deferred

Thanks for the well-formed proposal. The intelligence agent type idea is architecturally coherent — the shared-model / individualized-perspective pattern is a meaningful distinction from signals — and the seven open questions accurately map the design space the working group would need to resolve.

That said, this isn't in the current 3.1 or 4.0 build window, and a new top-level agent type in get_adcp_capabilities is a WG-level scoping decision. Parking in the spec backlog so it surfaces when agent-type expansion is on the agenda.

---

Triaged by Claude Code. Session: https://claude.ai/code/session_01K232MLtCNanUHj3CFhYx1p

---

Generated by Claude Code