Agentic AI companies sort into five layers, and working out which layer a vendor sits in tells you more than an hour on its website will. The homepages have converged on the same vocabulary. The engineering underneath them has not.
- Model providers, who sell the reasoning every layer above depends on.
- Framework and orchestration vendors, who sell the loop wrapped around that model.
- Agent development companies, who sell the integration labour.
- Vertical agent applications, who sell one finished job inside one industry.
- Infrastructure, covering identity and payment rails among others.
An agent, meaning a system that picks its own next step and calls tools without being told which one, is a narrow technical claim. A large share of what ships under the label is a fixed sequence of steps with a chat box bolted to the front. That is a useful product carrying a misleading name.
The sections below give you the map and then the test. The test matters more, because the map goes stale roughly every six months and nobody sends you the update.
AI agent vs agentic AI: what the companies are claiming
An AI agent is one system pursuing a goal with tools. Agentic AI is the umbrella term for the whole design approach, including setups where several agents pass work between them. Vendors use both words to mean whatever their product happens to do.
The distinction worth holding onto comes from Anthropic's engineering write-up on building effective agents, which separates workflows from agents. In a workflow, predefined code paths orchestrate the models and the tools. In an agent, the model directs its own process and its own tool use, keeping control over how the task gets done.
That line is the whole evaluation. Ask a vendor to name the decision their system makes that nobody wrote down in advance. If the answer is a branch in a flowchart, you are buying a workflow.
OpenAI's practical guide to building agents arrives at the same place from the other direction. It lists the components an agent needs:
- A model, doing the reasoning and the decision-making.
- Tools, meaning external functions or APIs the agent calls to take action.
- Instructions, the explicit guidelines and guardrails defining how it behaves.
The guide is blunt that applications integrating a model without letting it control execution, such as single-turn calls or sentiment classifiers, are not agents.
Both documents were written by companies with an obvious commercial interest in the word "agent" meaning something impressive. Both spend their opening pages arguing for a definition narrow enough to exclude most of what is being sold as one, which is either integrity or a very long game.
The five layers of the AI agent companies map
Every agentic AI company sells at one of five layers. The layer determines what you are actually buying, what it costs when it fails, and how quickly you can leave.
| Layer | What it sells | Switching cost |
|---|---|---|
| Model provider | Reasoning, tool-calling, context length | Low on paper, high in prompt tuning |
| Framework and orchestration | The agent loop, state, handoffs | Moderate, mostly rewriting glue code |
| Development company | Integration and delivery labour | High while the knowledge lives in their heads |
| Vertical application | A completed job in one domain | High, your data and workflow move in |
| Infrastructure | Identity, payment rails, evaluation, tracing | Low, these are the most replaceable pieces |
Most confusion in a buying process comes from comparing vendors at different layers. A framework vendor and a vertical application vendor can both answer "do you do customer support agents" with a yes, and mean completely different amounts of your engineering time.
The layer also predicts where a demo will be strong. Model providers demo raw capability. Development companies demo somebody else's finished outcome, which is a demo of their sales team.
Our breakdown of agentic AI use cases with public evidence behind them is the companion piece here. The map tells you who sells what; the use cases tell you which of it works yet.
AI agent frameworks and where the lock-in hides
AI agent frameworks give you the loop, meaning the code that calls the model and feeds each tool result back into it. Writing that loop yourself takes an afternoon. Writing the twelfth version of it, the one that survives retries and partial failures among others, takes a quarter.
Anthropic's guidance on frameworks is worth holding up against the sales pitch. Frameworks make it easy to start by simplifying low-level tasks, and they often add abstraction layers that obscure the underlying prompts and responses, which makes debugging harder. The recommendation is to use the model APIs directly first, then adopt a framework once you understand what it is hiding.
This is the advice teams skip and then rediscover at 2am, staring at a trace that shows the agent called the same tool nine times while the framework helpfully summarised all of it into one green tick.
When you evaluate a framework vendor, the question is what you can see. Ask to view the exact prompt sent to the model on a real run, including everything the framework injected. A vendor who cannot show you that string has built a product you will be unable to debug in production.
The second question is portability. Agent state and evaluation traces should be readable formats you can export. If the traces only exist inside their dashboard, your evaluation history is a hostage.
What an AI agent development company actually sells
An AI agent development company sells engineering hours wrapped in a fixed-price statement of work. The market is large, the quality range is enormous, and the good ones are distinguishable inside twenty minutes.
What they are really selling is integration, which is the honest description of where agent budgets go. The model is a small fraction of the system. The rest is connecting to a CRM whose API was last documented in 2019, then building the evaluation harness that tells you whether last week's prompt change helped. We went through that arithmetic in the AI systems engineering problem nobody budgets for.
Three things separate a competent agent development company from a repackaged web agency:
- They ask what a wrong answer costs before they ask what your budget is.
- They propose an evaluation set as a deliverable, with the test cases written down.
- They will name a task and tell you it should stay deterministic code.
That third one is the strongest signal available. A shop that has never talked a client out of an agent has never had a project fail in a way that taught them anything.
On AI agent development services more broadly, treat the pricing model as information. Fixed-price delivery on an agent build transfers the risk of an unpredictable system to whoever wrote the fixed price, and the way that risk gets managed is by narrowing scope until the agent is a workflow again.
How payment companies are building agentic AI commerce
The most consequential agentic AI companies at the infrastructure layer are the ones that already move money. Card networks and payment providers are building the rails that let a merchant tell an authorised agent apart from a fraud script.
The Agentic Commerce Protocol, an open standard released by Stripe and OpenAI under Apache 2.0, is the clearest published version of how this is meant to work. Payment credentials never reach the agent. A shared payment token passes between the parties instead, and the merchant keeps merchant-of-record status along with the right to reject a transaction by agent or by its own custom logic.
Read that acceptance-rules clause closely if you sit on the merchant side. It is the control surface for every agent-initiated purchase you will ever receive, and the defaults get set by whoever integrates first. Spending limits and per-agent rules belong in place before an agent has anything to spend.
The pattern generalises past payments. Identity and permission scoping are the infrastructure problems that decide whether agents are allowed near anything that matters, and they are being solved by companies that have handled untrusted actors for decades. Anyone who has spent a career watching banking standards move at geological speed can enjoy the sight of them sprinting.
Evaluating an agentic AI company, funding signals included
You can assess a vendor in one working session. Four questions do most of the work, and none of them require you to be technical.
- What decision does your system make that we did not specify in advance?
- Show me a trace of a real run, including the failed ones.
- What is in your evaluation set, and who wrote the test cases?
- What does it do when the same tool call fails three times in a row?
Question four is the one that separates products from demos. Every agent eventually hits a tool that returns an error, and a system with no answer for it will retry cheerfully until your rate limit intervenes.
On AI agent platform startup funding in 2026, the useful reading is directional. Capital has moved toward vertical agents and toward the control-and-security layer, which is a market saying that generic orchestration is becoming a commodity while domain depth and permission management are not. A vendor raising on distribution into one industry is making a more durable claim than one raising on a framework.
Funding is a signal about investor belief, so read it as evidence about survival odds. It says nothing about product quality. The practical version of the question: ask how long they can keep running the agent you are about to depend on.
Agentic AI Engineering takes these questions down to the tool and system-prompt level, which is where a vendor claim either holds or it does not. For the organisational side of the same decision, Agentic AI in Business covers autonomy boundaries and governance.
The one thing to carry out of all this: identify the layer first, apply the workflow-versus-agent test second, then ask for a trace. Vendors who welcome that third request are a small enough group that the exercise sorts the market for you, and our list of agents worth building for a business is a reasonable place to check what you could build instead.