Issue 04 — The AWS of agents will not be a model company

Last Wednesday afternoon I was wiring a new agent into V5’s code-quality pipeline. The fastest path in front of me was a six-line setup using a model provider’s native agent SDK. Typed. Documented. Batteries included. Good DX. I closed that tab and opened a blank file. I wrote a sixty-line harness against a model-neutral orchestration library instead.

That choice cost me two hours and most of a good coffee.

In 2029, one of those decisions is a rounding error and the other is a margin line on my P&L. I am betting — with my own time, every week — that the expensive choice is the cheap one.

This issue is about why.

PRED-004 — By December 2029, a dominant agent orchestration platform — a single player, or two to three in oligopoly — will control more than 30% of the agent orchestration market by revenue. It will be infrastructure-layer and model-agnostic. Not a model provider.

Confidence: 3 out of 5.

Start with the cleanest precedent the industry has.

In 2006 the cloud landscape was a mess. AWS, Google App Engine, Rackspace, GoGrid, Joyent, and a long tail of regional hosts all competed for the same “run my app somewhere” budget. No one was an obvious winner. By 2012, AWS was clearly dominant. By 2018, AWS, Azure, and Google Cloud together controlled about 67% of the market. Six years from fragmentation to a single dominant player. Twelve from fragmentation to a stable oligopoly.

The current agent orchestration landscape looks like cloud in 2008. LangChain, CrewAI, AutoGen, Semantic Kernel, Haystack, the Claude Agent SDK, the OpenAI Agents API. Overlapping primitives. No clear winner. Every month a new serious entrant. If the cloud consolidation pattern repeats — and infrastructure markets keep doing this — a dominant player crystallises by 2029.

That part is the easy half of the prediction. Here is the harder half, the part the chapter names but does not press on.

The winner will not be a model company.

Model providers have the strongest short-term incentive to own the orchestration layer. Bundle orchestration with the model, lock the developer in, capture the downstream economics. Every one of them is trying. Claude Agent SDK is model-locked by design. OpenAI Agents API is model-locked by design. Google’s agent stack is model-locked by design. These are not accidents. They are strategy.

Strategy loses here. The infrastructure layer rewards neutrality, and the history is relentless about this. AWS won cloud because it was not tied to an application. Stripe won payments because it was not tied to a bank. Cloudflare won the edge because it was not tied to a hoster. Every time the infrastructure winner was a vertically integrated incumbent trying to lock its own stack down, a neutral entrant ate it from the side.

So here is the test I actually use, on my own stack, every time I pick an orchestration library. Call it the let-go-first test: which of the current candidates can most credibly let go of its model preference first?

Run it down the list in April 2026. LangChain is mostly model-agnostic and has the distribution, but it is commercially fragile — the company has been churning through business model iterations. CrewAI and AutoGen are neutral by architecture but do not have the enterprise distribution. Semantic Kernel is Microsoft’s, which means it is model-neutral in theory and Azure-shaped in practice. Claude Agent SDK and the OpenAI Agents API are structurally incapable of letting go — the locked model is the product. So the race is really: which of the neutral-by-architecture players reaches commercial durability first, and which of the model-locked players lets go of the lock most convincingly when they realise they have to.

That is the race. The winner is whoever runs it best. And the practitioner receipt is that I am already betting on it: every orchestration decision I have made across my ventures since 2024 has refused model-locked SDKs, even when the DX advantage was obvious. Portability is the asset I am buying. In 2029 it is the asset the whole market will be buying.

The published falsification trigger:

If by December 2029, no player controls more than 15% market share, or if model providers own the orchestration layer as a bundled, inseparable part of their model offering, this prediction is wrong.

That is the stake as written. Here is the way I am actually most likely to lose it.

Not by having no winner. By having the winner be a model provider, and defining “inseparable” away. A 2029 world where the OpenAI Agents API controls 35% of orchestration revenue — because you cannot buy orchestration from them without a model contract — would be substantively my prediction being wrong while every dashboard shows it being right. The bundling is the hard case. If the model providers bundle aggressively enough and the neutral players do not reach distribution, the orchestration layer becomes a tax channel for whoever owns the model. That is the failure mode I am watching most closely.

Point me at a counter-example.

If you can name a model-provider-owned orchestration layer — OpenAI, Anthropic, Google, or a future entrant — that has crossed 30% share of agent orchestration revenue before a neutral, infrastructure-layer competitor has, reply to this email with the evidence. A specific number. A specific source. A specific quarter. I will publish the strongest counter-examples at atin-agarwal.com/predictions/pred-004-orchestration-platform-dominance/, with credit if the evidence holds.

Do the same in reverse. If you have deployment data from your own company showing your orchestration spend shifting away from model-locked SDKs toward neutral platforms over the last twelve months, send it. Both signals matter. The first tests whether I am wrong. The second tests whether I am early.

If you are a builder: do not marry a model-locked SDK for anything you expect to run in 2028. The productivity tax of a neutral orchestrator this year is smaller than the migration tax you will pay when you need to swap models and find the SDK was half your app. Your lock-in cost is invisible until you try to leave.

If you are a platform founder in this category: neutrality is not an engineering choice. It is your distribution story. Every buyer of agent infrastructure in 2027 will be asking how fast they can swap the model underneath. If your answer takes more than one sentence, you are not the winner.

If you are an investor: your filter at this layer is simple. “How fast could this company pivot models?” If the honest answer is “never without rewriting the product,” the company is a model-provider feature waiting to be repriced, not an infrastructure bet.

If you are a CTO doing vendor selection: write “let-go-first test” at the top of your evaluation rubric. Score each candidate on how credibly they could drop their current model preference inside a year. The scores will surprise you.

This issue is drawn from Chapter 9 of The AI Agent Economy — 15 falsifiable predictions with dates, numbers, and explicit triggers for being proven wrong. Pre-order on Kindle — $9.99. Release July 1, 2026. atin-agarwal.com/books

Read the full PRED-004 entry on the public tracking page → atin-agarwal.com/predictions/pred-004-orchestration-platform-dominance/

Previous issue: Issue 03 — The $100B market that doesn’t exist yet, and the invisible tax it will charge → SUBSTACK-003 Next issue: Issue 05 — 10,000 one-person conglomerates by 2028 — and the metric that tells you if you’re really one → SUBSTACK-005