An orchestration layer is the software that sits above a set of tools, services or artificial intelligence (AI) agents and coordinates them towards a single goal. It takes an objective, such as a campaign brief, breaks it into tasks, decides which system handles each one, keeps track of what has been done and checks the results against rules before anything is committed. The individual components do the work. The orchestration layer decides the order, carries the state between steps and holds the permissions.
The term is old in computing and new in advertising. It has spread through ad tech since late 2025 because agentic buying multiplies the parties in a transaction: a brand's agent, several sales agents, data providers, creative tools and the auction systems underneath. Without a coordinating layer, every agent would need a custom connection to every counterparty, and nobody would own the question of whether a purchase was allowed.
How the coordination works
Take a hypothetical four-week connected television (CTV) campaign with a budget of 200,000 euros. A buyer-side orchestration layer first converts the brief into structured objectives and constraints. It consults a registry to find sellers authorised to sell the relevant inventory, then sends the brief to several sales agents in parallel. Under the Ad Context Protocol (AdCP), each seller answers a get_products call with packaged products and prices. The layer scores the replies against budget, audience and brand-safety limits, allocates money across sellers and pauses for human approval. Only then does it issue create_media_buy calls, after which it reads delivery reports and shifts budget between sellers.
The documentation for AdCP version 2.5 draws this as a three-tier stack, with business principals at the top, technical execution at the bottom and an orchestration tier between them. Buyers and sellers, according to that documentation, "exchange impressions and money through the orchestration layer below." On the buy side, that tier holds a media orchestration platform that evaluates sellers and runs strategy. On the sell side, it holds sales agents and signal agents exposed as Model Context Protocol (MCP) servers.
IAB Tech Lab's Agentic Advertising Management Protocols (AAMP) turn the same idea into a hierarchy. Both its buyer and seller software development kits place orchestration at Level 1, as a Portfolio Manager, above channel specialists and functional agents. For buyers, that level handles budget allocation and performance management; for sellers, yield, deal acceptance and cross-selling. In the Tech Lab's open-source reference seller agent, the top level, called the Inventory Manager, runs on Anthropic's Claude Opus model while the lower levels use the smaller Claude Sonnet, according to its GitHub documentation. The reference buyer agent, built on the CrewAI framework, also assigns its orchestration level to Opus.
At impression time the word narrows. In the Agentic Real Time Framework (ARTF), the orchestrator is the host platform, usually a supply-side platform (SSP) or exchange. It calls partner containers over gRPC through a service named RTBExtensionPoint, passing an OpenRTB bid request and tmax, the millisecond budget for a reply. Containers return proposed changes, each tagged with a declared intent, and the orchestrator applies or rejects them. It is also expected to strip private and competitive signals before any container sees the request.
The tiers run on different clocks: weeks for a portfolio, hours for a campaign and, for an auction that must close in under 100 milliseconds, a few thousandths of a second.
From web services to agents
Orchestration entered software vocabulary through web services. IBM, Microsoft and BEA published BPEL4WS 1.0 in July 2002, a language for sequencing calls between services, and OASIS approved its successor, WS-BPEL 2.0, as a standard in April 2007. Containers borrowed the term next. Google announced Kubernetes in June 2014 and released version 1.0 on July 21, 2015, handing it to the newly formed Cloud Native Computing Foundation.
Language models gave the phrase its current sense. A Google whitepaper by Julia Wiesinger, Patrick Marlow and Vladimir Vuskovic, released in September 2024, split an agent into a model, tools and an orchestration layer, the last described as the cognitive architecture handling reasoning, planning and decisions. Anthropic's guide of December 19, 2024 set out an orchestrator-workers pattern, in which a central model breaks down a task, delegates the pieces to worker models and synthesises their results.
Marketing technology followed within a year. Adobe launched its Experience Platform Agent Orchestrator on September 10, 2025, built around a reasoning engine that interprets prompts and works out the steps needed. LiveRamp introduced agentic orchestration on October 1, 2025, letting outside agents reach its identity, activation and clean room tools through application programming interfaces (APIs). AdCP launched on October 15, 2025 with six founding companies. IAB Tech Lab opened ARTF, then called the Agentic RTB Framework, for public comment on November 13, 2025 and named AAMP on February 26, 2026. Its August 2026 finalisation notice dates the comment release to February 2026, while its standards page records comment closing on January 15, 2026.
Products carrying the name arrived in 2026. Magnite launched Magnite Orchestration on June 11, 2026 with dentsu and DIRECTV Advertising in beta, placing it above the transaction layer. The same day Mediaocean introduced NIVO, and Zvika Netter, its chief innovation officer, summarised the division of labour: "NIVO thinks, Orchestrator connects, and Agents act." ARTF 1.0 was declared final on August 12, 2026. According to IAB Tech Lab, the final text added an orchestration layer letting one node call nodes run by other organisations and merge their answers into a single change to the bid request.
Why the layer matters to marketers
Whoever runs the orchestration layer decides which sellers are asked, what data travels and what gets approved, so fees and accountability gather there. Magnite's chief executive, Michael Barrett, told tipsheet in July 2026 that the layer "also handles settlement, policy enforcement, pricing, yield management, identity, payments and compliance." He likened it to the clearing function of a stock exchange. In the same interview he put the most optimistic 2027 forecasts for agentic spend at $600 million to $700 million, small against programmatic as a whole.
Integration cost explains the interest. Adform's chief technology officer, Jochen Schlosser, estimated custom SSP deal integrations at 250 to 350 days of work, against 50 on a shared standard. A layer that speaks one protocol to many sellers turns that gap into a product. Agencies have built their own: an executive at Omnicom's PHD described Omni as the holding company's marketing orchestration platform, with all of its agentic tools running inside it.
Costs, control and open disputes
The definition itself is contested. Google's framework puts reasoning inside the orchestration layer. Mediaocean treats orchestration as connective plumbing beneath a separate intelligence layer. ARTF applies the word to a host server accepting or rejecting changes in milliseconds. Products sharing the label do different jobs.
Neutrality is a second question. Magnite presents Orchestration as neutral, yet it operates the coordination layer, the Seller Agent and the Buyer Agent, and its September statement did not say which protocol a connecting agent must speak. The standards compete too. Shailley Singh, IAB Tech Lab's chief operating officer, counted 13 functions on which AAMP and AdCP overlap. AdCP defines 21 specialisms for providers but none for the orchestrators calling them, according to a request for comments posted to its GitHub repository in May 2026.
Cost is the third. Every call an orchestrator chains consumes model tokens. Draft Digital, a Dutch operator, collapsed twelve MCP calls into a single buyer agent because twelve endpoints meant twelve context loads and twelve authentication paths. Evidence on results is thin. DataBeat found agentic buyers clearing at $6.13 CPM against $6.95 for conventional demand, a 13.4% gap, while entering 86% fewer auctions. The data measures price paid, not business outcome.
Access is the fourth. Google's developer policies of August 31, 2026 prohibit programmatic proxies, naming MCP servers among intermediate layers that solely re-expose Google Ads capabilities. A third-party orchestrator can be cut off by the platform it coordinates. ARTF raises the opposite worry, since decisioning runs inside a sealed container that neither party inspects. Governance is the industry's answer so far: PubMatic's guardrails of August 5, 2026 add approval workflows authenticated by one-time passcode and audit trails.
Not the same as
Mediation and header bidding. Both decide which demand source fills a single impression. Header bidding calls bidders together; mediation is the app equivalent. Neither plans or books a campaign.
Journey orchestration. Customer engagement platforms use the phrase for sequencing messages to one person. Iterable runs it through Journeys, renamed from Workflows.
Container orchestration. Kubernetes schedules compute across servers. ARTF runs on such infrastructure, but that orchestration concerns machines, not media decisions.
Protocols. MCP, AdCP and AAMP are shared languages. An orchestration layer is the software that speaks them.
Recent developments
Magnite and ITN said on September 10, 2026 that ITN's seller agent would sit inside Orchestration for local linear television, with full integration due in the fourth quarter. Six days later Magnite extended the layer to linear and addressable TV, home screens and pause ads, audio, and display and online video, and opened it to software vendors such as Newton Research. It disclosed no volumes or prices.
Standards work continues. IAB Tech Lab published AAMP 3.0 on September 22, 2026, introducing OpenProposal, a draft specification for how buyer and seller agents exchange proposals before a commitment. Public comment runs until October 22, 2026. Proposals for ARTF 2.0 are open. As of September 2026, the layer exists in competing versions, and the contest is over which one buyers and sellers will trust with money.
Timeline
- July 2002: IBM, Microsoft and BEA publish BPEL4WS 1.0, an orchestration language for web services
- April 2007: OASIS approves WS-BPEL 2.0 as a standard
- June 2014: Google announces Kubernetes
- July 21, 2015: Kubernetes 1.0 is released and the Cloud Native Computing Foundation is formed
- September 2024: Google's Agents whitepaper names the orchestration layer as one of three agent components
- November 2024: Anthropic introduces the Model Context Protocol
- December 19, 2024: Anthropic publishes Building effective agents, describing the orchestrator-workers pattern
- September 10, 2025: Adobe launches Experience Platform Agent Orchestrator
- October 1, 2025: LiveRamp introduces agentic orchestration
- October 15, 2025: The Ad Context Protocol launches with six founding companies
- November 13, 2025: IAB Tech Lab opens the Agentic RTB Framework for public comment
- January 5, 2026: PubMatic launches AgenticOS
- January 15, 2026: The ARTF public comment period closes
- February 26, 2026: IAB Tech Lab names AAMP, with orchestration at Level 1 of its buyer and seller hierarchies
- June 11, 2026: Magnite launches Magnite Orchestration; Mediaocean launches NIVO
- July 16, 2026: tipsheet publishes Michael Barrett's interview on the orchestration layer
- August 5, 2026: PubMatic launches guardrails on AgenticOS
- August 12, 2026: ARTF 1.0 is declared final, adding a cross-node orchestration layer
- August 20, 2026: IAB Tech Lab counts 13 functions shared by AAMP and AdCP
- August 31, 2026: Google prohibits programmatic proxies, naming MCP servers
- September 10, 2026: ITN's seller agent joins Magnite Orchestration
- September 16, 2026: Magnite extends Orchestration to linear TV, home screens, audio and display
- September 22, 2026: IAB Tech Lab publishes AAMP 3.0 and OpenProposal for comment
- October 22, 2026: The AAMP 3.0 comment period is scheduled to close
Related PPC Land coverage
- IAB Tech Lab names its agentic ad initiative AAMP to end market confusion - The three pillars and the Level 1 Portfolio Manager in AAMP's buyer and seller hierarchies.
- Explaining ARTF - How the host platform acts as orchestrator for containers inside the auction.
- AI Agents: Google unveils framework for next-gen systems - The September 2024 whitepaper separating model, tools and orchestration layer.
- Adobe launches AI agents for enterprise customer experience management - The September 2025 Agent Orchestrator and its reasoning engine.
- LiveRamp introduces agentic AI tools for marketing automation - Agentic orchestration giving outside agents API access to identity and clean room tools.
- Ad Context Protocol (AdCP) launches for advertising automation - The October 2025 launch of the protocol that defines an orchestration tier.
- Magnite Orchestration wants to be the wiring behind AI ad buying - The June 2026 launch of a sell-side coordination layer for buyer agents.
- Mediaocean's NIVO AI cuts campaign setup by 90% with Innovid agents - A buy-side split between intelligence, orchestration and execution, with vendor-reported pilot figures.
- Magnite CEO caps 2027 agentic ad spend near $700m as AI cuts contractor tasks - Michael Barrett on what an orchestration layer must settle and how much agentic spend to expect.
- Neither of ad tech's two AI agent protocols runs at scale, Adform CTO says - Jochen Schlosser's estimate of custom versus standard integration effort.
- Omnicom's PHD gains just 10-20% productivity from AI in APAC - Omni as a holding company orchestration platform and the productivity it has delivered so far.
- AI buying agents gain Samsung, iHeartMedia and DIRECTV ad space via Magnite - The September 2026 expansion and the limits of Magnite's neutrality claim.
- AAMP and AdCP overlap on 13 functions, IAB Tech Lab COO says - Shailley Singh's rebuttal of the idea that the two frameworks occupy separate layers.
- Draft Digital cuts 12 MCP calls to one buyer agent to stop token burn - How stacked agent connections multiply token costs.
- Programmatic buyers gain 13.4% CPM edge over AI agents, DataBeat finds - Clearing prices and auction participation for agentic and conventional demand.
- Google bans programmatic proxies from Ads API access - The August 2026 policy restricting third-party intermediate layers, MCP servers included.
- PubMatic guardrails block AI agents from unapproved ad inventory - A five-step governance framework constraining what agents may transact.
- Explaining header bidding - The auction technique often confused with orchestration, and its relationship to mediation.
- Explaining Iterable - Journey orchestration in a customer engagement platform.
- Magnite and ITN cut local linear TV buying from weeks to hours with AI agents - A broadcaster's seller agent connected to Magnite's coordination layer.
- IAB Tech Lab gives industry until October 22 to weigh in on AAMP 3.0 - OpenProposal and the draft status of AAMP 3.0.
Summary
Who. Standards bodies and protocol groups, chiefly IAB Tech Lab and AgenticAdvertising.org, which publishes AdCP; sell-side platforms such as Magnite and PubMatic; buy-side and data companies including Mediaocean, LiveRamp and Adform; holding companies such as Omnicom; and enterprise software vendors such as Adobe. Model providers including Google and Anthropic supplied the vocabulary.
What. The software tier that coordinates tools, services and AI agents towards a goal: breaking briefs into tasks, routing them, holding state and permissions, and checking results before money moves. In advertising it appears as AdCP's orchestration tier, AAMP's Level 1 Portfolio Manager, the ARTF host orchestrator and branded products such as Magnite Orchestration.
When. The idea dates to web service orchestration in 2002 and container orchestration from 2014. Its agentic meaning took shape in 2024, reached advertising between September and November 2025, and moved into commercial products and final specifications through 2026, with AAMP 3.0 open for comment until October 22, 2026.
Where. Across the transaction: in buyer and seller agent systems for planning and deal-making, inside SSPs and exchanges at impression time, and in marketing clouds and agency platforms, largely in the United States and Europe.
Why. Agentic buying adds parties to every transaction, and something has to sequence their work, enforce budgets and approvals, and settle what was bought. The layer that does so gathers control over data, fees and accountability, which is why vendors, standards bodies and platforms are contesting its definition, its neutrality and who may connect to it.
Discussion