Stop renting AI agents. Own the ones that make you different.

Every software vendor your company already pays now has an agent to sell you. The CRM has one, the helpdesk has one, the HR platform has one, and a new wave of start-ups promises one agent that will run your whole workflow end to end. The demos are good. The pitch is the same everywhere: switch it on, point it at your data, and the work does itself. Then the pricing arrives — per seat, per conversation, per resolution, per “agent action” — and a team that wanted to automate three annoying processes finds itself negotiating an enterprise contract.

We think most companies are about to overpay for AI agents in the same way they overpaid for the last generation of SaaS, and for the same reason: they are renting a generic answer to a problem that is specific to them. The better move for a lot of teams is to own the agents that matter — a small library of agents and workflows built around how your company actually works, running on models you can swap, wired into the systems you already have. It costs less than people assume, and it is the only version that is genuinely yours.

Why vendor agents get expensive

The cost problem is not that vendors are greedy. It is structural, and it shows up in three ways.

• You pay a margin on the model. Underneath almost every vendor agent is a frontier model you could call directly. The vendor adds a platform, a pricing unit, and a margin on top. That is fair for the convenience — but the unit it charges by tends to scale with exactly the thing you wanted more of. The more the agent does, the more it costs, which quietly punishes success.
• You pay once per system. The CRM agent only knows the CRM. The helpdesk agent only knows the helpdesk. The workflow that actually matters to your team crosses four systems, so you end up licensing four agents that cannot see each other, then paying someone to glue them together.
• You pay for the generic version. A vendor agent is built for the median customer. Your pricing rules, your approval chain, the exception your largest client negotiated three years ago — none of that is in the product. So the team works around the agent, or pays for professional services to customise something that was never designed to be customised that far.

And then there is lock-in. Once your prompts, your workflow logic, and a year of tuning live inside someone else’s platform, the renewal conversation is not really a negotiation. Moving means rebuilding from scratch, and the vendor knows it.

What “your own agent library” actually means

This is where the idea usually gets misheard as “build your own AI platform”, which sounds like a multi-year programme with a large team. It is not that. The model providers have done the hard, expensive part. What is left is a thin layer that is specific to you, and that layer is small enough to own.

• Tools. Small, well-tested connectors that let an agent read from and act on your systems — look up a customer, draft a quote, open a ticket, update a record — each with its own narrow permissions. You write each one once and every agent reuses it.
• Skills. Your company’s know-how, written down in a form an agent can follow: how you qualify a lead, how you triage a support case, how finance handles a disputed invoice. This is the part no vendor can sell you, because it is the part that makes you different.
• Workflows. The sequences that tie tools and skills together for a specific team, with the review steps and escalation rules that team needs. Sales gets one shape, operations another, finance a third — built from the same parts.
• Evals. A standing test set per workflow that defines what a correct result looks like, so you can change a prompt or switch a model and know within minutes whether anything got worse.
• A model layer you can swap. The agents call models through one routing layer, so a cheap, fast model handles the routine cases and a stronger one takes the hard ones — and when a better or cheaper model ships, you change a setting rather than a contract.

Put together, that is a library: reusable pieces your teams can compose into new agents without starting from zero each time. The second agent is much cheaper than the first, and the fifth is cheaper still, because most of what it needs already exists. That compounding is the opposite of the vendor model, where every new use case is a new line item.

How we build it without it costing millions

The expensive way to do this is to start with the platform: design the grand architecture, build the framework, and hope the use cases arrive. We do it the other way round. We start with one workflow that a real team does every day, where the volume is high enough to matter and a mistake is recoverable. We build that one agent properly — tools, skills, review gates, evals, observability — and put it into production. The reusable library is what falls out of doing the first few well, not something designed up front.

That keeps the spend tied to results. Each agent has to earn its place against a measurable outcome before the next one starts. It runs on the cloud and the SDKs your engineers already use, so there is no new platform to license and no proprietary runtime to be locked into. And the code, the prompts, the skills, and the evals all live in your repository. This is the heart of an AI Agents engagement with us: we build the first agents alongside your team, shape the shared library as we go, and hand the whole thing over with the runbooks your engineers need to extend it themselves. The goal is that you stop needing us, not that you start needing us for everything.

Owning it also changes the economics of the model bill. When you call the models directly, you pay the provider’s price and nothing else, you route routine work to cheaper models, and you can see exactly what each workflow costs to run. Cost becomes something you engineer, not something you are invoiced.

The honest counter-argument

Building is not free, and it is not always the right answer. Owning an agent library means owning its maintenance: someone has to keep the connectors working when an upstream API changes, keep the evals current, and respond when an agent misbehaves. A vendor absorbs that for you, and for a small team with no engineers that is worth a great deal. Vendors also ship features you would never build yourself, and for genuinely commodity work — transcribing calls, summarising tickets, drafting routine emails inside the tool you already use — the built-in agent is often good enough and cheaper than anything custom.

So the real rule is not “build everything”. It is: buy the commodity, build what differentiates you. If a workflow is the same at every company in your industry, rent the vendor’s agent and move on. If a workflow encodes how you win — your sales motion, your service standard, your pricing judgement — that is the part worth owning, because it is the part a generic product will always get slightly wrong. Most companies find their differentiating workflows are a short list. That is exactly why owning them is affordable.

What to do in the next 30 days

• List every AI agent you pay for or are being pitched. For each, write down what it costs at your expected volume in a year, not at pilot volume.
• Sort your workflows into commodity and differentiating. Be honest. The differentiating list should be short.
• Pick one differentiating workflow and scope an owned agent for it. High volume, clearly definable, recoverable when wrong. Write the eval set first.
• Check your exit costs. For any vendor agent you keep, ask where your prompts, workflow logic, and tuning live, and what it would take to move them.

The market will keep producing agent products, and some of them will be excellent. But the companies that get the most from AI will not be the ones with the most agent subscriptions. They will be the ones that own the handful of agents that encode how they work — built on models they can swap, in code they control, getting cheaper with every one they add. Renting generic intelligence is easy. Owning the part that makes you different is where the value is.

Being pitched agent platforms and worried about the bill — or the lock-in? Talk to Cravings about building your own agent library. We build your first production agents around the workflows that make you different, in your repository, on models you can swap, and hand your team a library it can keep extending without us.