When the Buyer's Agent Meets the Seller's Agent
A natural-language brief in Claude became live CTV spend through a protocol handshake. Ten surfaces are becoming agentic counterparties — but autonomy and interoperability are inversely correlated, so far, and the consolidation ledger shows exactly what agents are hungry for.
In English, please
On December 15, 2025, someone typed a plain-English ad campaign brief into Claude. The request went to software representing the advertiser, which had a structured back-and-forth with software representing the ad seller -- two programs following an agreed script so they could propose, check, and confirm a deal without a person relaying every step. The exchange actually closed: real money bought real streaming-TV ad time, for a hard-seltzer brand called Clubtails, run through ad-tech company PubMatic and agency Butler/Till.
The vendor reported the deal ran 87% faster to set up, 70% faster to troubleshoot, and 30% cheaper per ad shown. Those numbers come from the vendor, not an independent audit, and from a showcase where the vendor's own engineers were optimizing in real time. The case proves two AI systems can strike a deal this way -- not that it reliably saves money once nobody is standing by to help.
Almost everywhere else, that kind of automated negotiation is still on a leash. Most "agentic AI" tools that ad-buying platforms and marketing software makers are showing off stop short of letting the AI spend money on its own -- a person still has to approve the final step. Most of what gets marketed as an autonomous AI agent today is really a fast assistant with a human keeping a finger on the button.
There is also a split in how companies are building this. Large single-company platforms, like Meta and Amazon, are moving fastest toward letting AI act with less human involvement -- but only for buying ads on their own turf, in ways outsiders can't easily audit. Efforts to build a shared, open system where any company's AI can negotiate with any other's -- the approach the Clubtails deal used -- are moving slower and keeping more human sign-off, because they have to work across companies instead of inside just one.
Big ad and marketing companies are also on a buying spree for two kinds of businesses: those holding customer identity data (who a customer actually is across platforms), and those making the software that decides which ad to show which person. Automated buying and selling needs both ingredients, so companies are stocking up before this new automated marketplace has proven it works reliably.
On this page
From Brief to Buy: The First Agentic Handshake · The Leash Audit: “Agentic” in the Deck, Sign-Off in the Contract · Walled Autonomy Versus the Open Mesh · Ten Surfaces, One Question · The Consolidation Ledger: Fuel and Steering · Monday Morning · Appendix: the receipts
A four-part series — Part 1: The CDP grew agents · Part 2: The customer grew agents first · Part 3: The market grew agents on both sides (you are here) · Part 4: The CMO owns the action space
| Term | Definition |
|---|---|
| Action space | The bounded set of goals, content, audiences, budget rules, policy checks, approvals, measurement gates, and execution surfaces an agent is allowed to act inside. |
| Agentic marketing | Marketing where software does not just recommend actions, but selects, sequences, negotiates, or executes them inside human-defined constraints. |
| CMO role | Not to approve every action, but to design the constraints and proof system that govern agent behavior. |
On December 15, 2025, someone typed a natural-language campaign brief into Claude. A buyer agent parsed it, opened a protocol negotiation, and a seller agent on the other end answered. What came back was not a slide or a simulation — it was real spend on live CTV inventory, for a hard-seltzer brand called Clubtails, run by PubMatic with agency Butler/Till. The claimed results: 87% faster campaign setup, 70% faster troubleshooting, 30% lower CPMs [Vendor-claimed]. Hold the caveat as tightly as the headline: this was a controlled showcase with PubMatic engineers optimizing in real time. It proves feasibility, not economics.
But feasibility was the missing piece. Part 1 of this series watched the CDP grow agents in a single business week. Part 2 found the customer’s agent already deployed at near-billion scale before any CMO had one. This part is the market itself: the moment the buyer’s agent and the seller’s agent met and closed. Surfaces stop being channels and become endpoints on a negotiation protocol — campaigns do not disappear in that world; they stop being the primary operating unit — and the question at every endpoint is the same: who negotiates here — your agent, their agent, or nobody’s yet? (Tags in this part follow the Part 1 legend: [Confirmed] / [Vendor-claimed] / [Reported] / [Author inference].)
From Brief to Buy: The First Agentic Handshake
The handshake runs on AdCP — the Ad Context Protocol, built on Anthropic’s MCP, launched October 15, 2025 with founding members Scope3, Yahoo, PubMatic, Swivel, Triton Digital, and Optable. The arc since is one of the fastest visible protocol-to-live-media runs advertising has produced: first autonomous CTV campaign in December 2025, first agentic OOH campaign in May 2026 with human oversight throughout (participants in the receipts) — launch to two live mediums in under eight months, with audio seated at the founding table from day one.
Disclosure, stated plainly: this author holds a seat on AdCP’s Signals & Measurement working group. And “over 100 members” is AdCP’s own claim, not an audited count [Vendor-claimed].
The protocol layer comes with an identity layer — B2A, brand-to-agent. In plain English: B2A readiness means making the brand machine-readable enough that an agent can verify who you are, what you sell, what claims you authorize, and which policies govern transactions. AdCP’s docs already specify brand.json at /.well-known/brand.json with a verify_brand_claim tool [Confirmed].
Contrast the cargo cult that preceded it: llms.txt. Ahrefs found that 97% of published llms.txt files received zero requests [Confirmed, Ahrefs study; full ratios in the receipts]. Marketers performing B2A readiness rather than achieving it — a caution that applies to whichever badge is fashionable next, including brand.json, until consumption is measured.
One bridge back to Part 1: the flagship agentic CDP launch never mentioned this protocol layer once — and that interop void is where orchestration of the next decade actually gets decided.
The Leash Audit: “Agentic” in the Deck, Sign-Off in the Contract
The handshake is real. The second finding is that almost everywhere else, autonomy is leashed. Run the audit against the vendor decks and one pattern falls out — copilot with a signature line.
| Vendor | Agent | The leash |
|---|---|---|
| The Trade Desk | Koa Agents (Apr 2026, alpha) | Autonomous buying explicitly off-limits |
| Kovva | Buyer agent | Automates everything short of the transaction, by design |
| Stagwell | The Media Machine | Humans at every decision gate |
| Ask Advisor (beta) | Acts with oversight | |
| Adobe | AEP agents + Orchestrator | ”Autonomous WITH human oversight” |
| Klaviyo | Composer | ”Nothing goes live without approval” |
Verified public examples of autonomous or policy-bounded spend do exist — Clubtails, above; Scope3’s policy-bounded transactions with fees public on its pricing page [Confirmed]; Magnite’s buyer-agent pilots; and PubMatic’s later claim of 30+ fully autonomous end-to-end agentic campaigns by Q1 2026 [Vendor-claimed, via AdExchanger] — the full caveat set for each is in the receipts. The list grows monthly, which is exactly why it is a list of examples and not a census.
The market, meanwhile, votes with contracts, not demos: The Trade Desk signed 45 joint business deals in March 2026, up 55% year over year — while keeping the agent leashed.
And watch the fee stack, because agentic disintermediation can become agentic re-intermediation:
| Old fee | New agentic version |
|---|---|
| DSP fee | Buyer-agent execution fee |
| SSP fee | Seller-agent negotiation fee |
| Agency fee | Agent orchestration fee |
| Verification fee | Permission / policy-check fee |
| Measurement fee | Referee / decision-gate fee |
The tell is whether agent fees displace the DSP-SSP-agency chain or simply stack on top of it — publishers are already lodging the complaint (on record, in the receipts).
The structural finding, carried into every vendor meeting: across the public claims checked here, I found no independent public audit of agentic-marketing lift. Every ROI number in this part is vendor-supplied. You are being asked to fund the category on case studies the vendors wrote.
Walled Autonomy Versus the Open Mesh
Here is the strategic spine of this part, and arguably of the whole series: autonomy and interoperability are inversely correlated — so far. Full automation is arriving fastest exactly where it is least portable, while the interoperable systems stay slower and supervised.
The walled evidence first. Meta’s stated end-state — objective plus bank account in, everything else automated, by end of 2026 [Reported, WSJ-sourced] — is the closest thing to true autonomy shipping at scale, and it works only inside Meta.
Amazon runs the same play with a lawsuit attached: its own Ads Agent and MCP server welcome agent-mediated demand — through Amazon’s agents — while Perplexity’s Comet gets sued off the shelf.
The fork, as a table:
| Camp | Who | The trade |
|---|---|---|
| Walled autonomy | Meta, Amazon, Google, Netflix | Full automation, low portability, opaque measurement |
| Open mesh | AdCP, Scope3, PubMatic, Magnite (via pilots), Triton, identity/data clouds | Interoperable, slower, more supervised |
| Enterprise agentic CDP | Databricks, Hightouch, Salesforce, Adobe, Braze, MoEngage | Owns decisioning, may not own media execution |
The historical rhyme is exact: the open-protocol world and the walled-garden world are re-running the composable-versus-packaged argument one layer up — the central tension of this series’ prequel, promoted. And the CMO’s fork is genuinely uncomfortable: the gardens sell autonomy you cannot audit; the mesh sells auditability that is not yet autonomous. Neither side resolves it for you — the CMO must manage both worlds, and the action space, Part 4’s subject, has to span them.
One paragraph on the creative layer, because the whiplash prices the risk. Coca-Cola’s AI holiday ad cut positive sentiment from 23.8% to 10.2%, and Coke doubled down anyway for World Cup 2026 [Reported] — while Icon, the Founders Fund-backed “first AI CMO,” now sells human-made ad packages as “The Human Admaker” (pricing in the receipts). Production economics and audience sentiment are pulling in opposite directions, and nobody has priced the brand cost of losing that argument.
Ten Surfaces, One Question
Walk the surfaces one by one.
- Email: the inbox grew an agent on both ends — Gemini-era inbox features reading marketing before recipients do (per Part 2’s downgraded but directionally intact claim), and Klaviyo’s 196,000+ customers getting prompt-to-campaign generation.
- Search/LLM: Pew’s 8%-versus-15% click rate is the surface’s epitaph for the click.
- Social: furthest along, most closed — Meta’s $55.0B quarter, up 33% year over year [Confirmed, Q1 2026 8-K].
- CTV: the first real autonomous campaign happened here, and Netflix is testing — a test, not a deployment — agents that manage, optimize, and purchase its ads (scale estimates in the receipts).
- Retail media: Amazon is the paradigm — Rufus retired into Alexa for Shopping, an Ads Agent with directional cost reductions [Vendor-claimed; numbers in the receipts], and an MCP server for external agents.
- Commerce: Google’s Universal Cart spans Search, Gemini, YouTube, and Gmail — when the cart follows the user across agent surfaces, “channel” stops being a planning unit.
- The rest of the tour — display, cinema, OOH, audio — is in the receipts below.
Now score the same map by who actually holds the agent at each surface:
| Surface | Buyer agent | Seller agent | Platform agent | Customer agent |
|---|---|---|---|---|
| Search / LLM | Low | Low | High | High |
| CTV | Emerging | Emerging | Medium | Low |
| Retail media | Low | Medium | High | Medium |
| OOH | Emerging | Emerging | Medium | Low |
Stop asking “what’s our channel mix.” Start asking, per surface: who negotiates here — our agent, their agent, or nobody yet?
The Consolidation Ledger: Fuel and Steering
Read the 2025–26 acquisition ledger with one lens and it stops being noise: every acquirer is buying the same two things — identity substrates and decisioning engines — because agents need fuel and steering.
The identity column has one thesis deal: Publicis–LiveRamp, $2.546B equity value, $2.167B enterprise value, close expected H2 2026 [Confirmed, Publicis PR] — under a press-release title that literally says “to accelerate data co-creation for smarter agents.” Publicis ↗
The decisioning-and-measurement column: Braze acquired OfferFit for $325M, completed June 2, 2025 [Confirmed, Braze Investor Relations]. MoEngage bought Aampe. BlueConic bought Blueshift. Sitecore bought Scrunch. Broadsign bought Place Exchange. Notice what the last three have in common: the measurement and context layers are being bought before they are validated — acquirers paying for position in a category that is, per the audit above, still pre-audit.
Monday Morning
| From | Do this |
|---|---|
| Part 1 | Audit where agent memory, campaign state, and Golden Context will live. |
| Part 2 | Stop treating AI answer visibility as SEO. Treat it as demand interception. |
| Part 3 (this part) | Ask which surfaces are protocol-ready, copilot-only, or truly autonomous. |
| Part 4 | Add agent-state ownership, native holdouts, and decision logs to procurement. |
Appendix: the receipts
Ten surfaces, the full tour. Display/programmatic: copilots everywhere; autonomy confined to the supervised pockets in the leash audit. Cinema: not an agentic buying surface yet — the proving ground for prompt-based prediction via Emberos’s Wicked: For Good case [Vendor-claimed]. OOH: the most surprising first mover — the Dutch handshake ran with Broadsign on the sell side, Draft Digital buying, and Global Netherlands as the advertiser, on a supply side consolidated to 1.8M programmatic screens (Broadsign + Place Exchange), after T-Mobile’s $600M all-cash Vistar close in February 2025, a pre-wave consolidation. Audio: protocol-ready, deal-flow pending — Triton sat at AdCP’s founding table, but no public agentic audio campaign exists as of July 2026; the surface’s entire verified agentic footprint is a seat at the table. Social depth: Meta reports 8M+ advertisers using genAI creative. CTV depth: Netflix’s test spans 250M+ ad-tier viewers and a ~$3B ad business [analyst/secondary estimate, not an audited Netflix number]. Amazon depth: the Ads Agent claims directional ~16% CPA and ~18% CPM reductions [Vendor-claimed, unofficial]; Rufus reaches 300M+ customers on a ~$10B incremental-revenue track per Jassy [Reported]; Amazon tells advertisers to expect meaningful demand to resolve inside agent answers rather than on the results page.
The llms.txt ratios, in full. Ahrefs studied 137,210 domains: 28% publish an llms.txt, 97% of those files received zero requests, and 96% of the remaining traffic was bots [Confirmed, Ahrefs study].
Autonomy examples, full caveat set. Scope3: policies checked before every execution; bounded autonomy that is never marketed as “independent negotiation” [Vendor-described]. Magnite: buyer-agent pilots with Kepler and MiQ on Disney inventory — pilots, not production.
Consolidation, non-thesis balance. Omnicom closed IPG in November 2025 — $25B+ pro forma revenue, roughly 4,000 roles cut [Reported], Acxiom Real ID and Flywheel wired into Omni, stated plan to disintermediate ad-tech middlemen; the deal value varies by the date you check, so none is cited. Uniphore–ActionIQ (~$145M, December 2024). Sitecore–Scrunch: ~$225M [Reported]. Post-Publicis-LiveRamp, the identity marketplace inside the “neutral” lakehouse is largely holdco-owned — Acxiom via Omnicom/IPG, LiveRamp via Publicis, pending — which is why the data cloud auditions as neutral venue while the lakehouse handles first-party matching in place [Author inference].
Clubtails, full caveat set. Controlled showcase; PubMatic engineers optimizing in real time; all results vendor-claimed; feasibility proven, economics not.
The PubMatic Q1 2026 claim, in full. 30+ fully autonomous end-to-end agentic campaigns and 1,000+ direct publisher deals [Vendor-claimed, AdExchanger] — the number that retired this series’ earlier “exactly three pockets” framing. Pockets multiply; the framing that survives is examples-not-census.
Fee-stack detail. Scope3: 4% on routed, 8% on decisioned transactions [Confirmed, public pricing page]. Scope3’s in-media margins land on top of DSP, SSP, and agency fees. Publisher complaint on record: “yet another fee.” The replace-or-join test: if agent fees stack on DSP + SSP + agency fees rather than displacing them, agentic trading has re-intermediated the very chain it promised to collapse.
Creative-whiplash receipts. Icon: Founders Fund-backed, $9.2M raised, pivoted to “The Human Admaker” at $399–$999. Coke sentiment: 23.8% → 10.2% on the AI holiday ad [Reported]. Roughly a third of video commercials already use AI tools [Reported, 2026].
The buyer grew agents; so did the seller. The gardens went autonomous and opaque; the mesh went interoperable and supervised. Which means the last scarce thing in marketing is the referee — and the CMO who wants to keep steering has to own the rules, not the campaigns. Part 4: the operating model, the ownership split, and the ten questions to ask every agentic vendor before you sign.