The agent is the easy part
Two platform companies shipped agent platforms the same week and said the quiet part: the agent loop is one percent of the work.
In one week this June, Vercel and Databricks, a frontend infrastructure company and a data company with nothing obvious in common, both shipped full platforms for building AI agents. Vercel released an agent stack, an open-source framework, an infrastructure agent, and enterprise controls. Databricks expanded Agent Bricks into a platform that has already built over a hundred thousand agents and processes more than a quadrillion tokens a year. The numbers are real and the engineering is serious.
But the most useful sentence in either announcement was not a feature. It was Databricks describing the shape of the problem: the core agent loop is just one percent of the work, and the other ninety-nine percent is the hidden technical debt of agentic systems, the token capacity, the deployment, the security, the evaluation, the monitoring, the context, and the sharing.
Sit with that. Two companies that just spent enormous effort building agent platforms are telling you the agent is the easy part. The thing the demos show you, the clever loop that reads your inbox and drafts a reply, is one percent. Everything that makes it safe, accountable, affordable, and correct in production is the other ninety-nine, and that ninety-nine is invisible in every viral clip.
Why this is the only number that matters
The agent loop is a commodity now. You can get a capable one from Anthropic, from OpenAI, from an open-source model, from a framework you install in an afternoon. That part is solved well enough that two platform companies treat it as a rounding error. If the loop were the product, nobody would need a platform around it.
So the value is not the model and it is not the loop. The value is the ninety-nine percent: did the agent only touch what it was scoped to touch, can you prove what it did after the fact, did it stay inside budget, did it ground its answer in something real or did it make something up, and is there a human between it deciding and it acting. Those are not features you bolt on at the end. They are the entire job, and they are exactly the work that does not film well.
We have been saying a version of this for a while in plainer words: the model is a commodity, and the trust layer is the product. It is a better feeling to hear it confirmed by a company doing a quadrillion tokens a year than to keep asserting it ourselves.
What we build, and what we deliberately do not
We build the ninety-nine percent, scoped small, on purpose.
We are not building a horizontal agent platform. That race now has Vercel and Databricks in it, and it is the platform we chose not to chase. Our lane is the focused agent and the trust layer around it, sized for businesses that are not the Fortune 100 and will never run a quadrillion tokens.
Inside that lane, the commitments are the ninety-nine percent made concrete. Our security agents run inside the customer's own cloud, with the data never leaving, where the big platforms offer that as an enterprise upgrade. They read by default and gate every write behind a separate token, a dry run, a live re-check, and automatic rollback. They ground every claim in the real resource they inspected, and a claim that cannot trace to something real is reported as a failure, not shipped. There is a permanent ceiling no agent crosses on its own, and a human sits between deciding and doing. When one of the platforms ships an infrastructure agent that investigates an issue and then recommends a fix for human approval, we read it as agreement: that gate is the right instinct, and it is one we already hold.
The honest framing is that these platforms are mostly a tailwind for us. They are commoditizing the plumbing we would otherwise have to build, the gateways and sandboxes and autoscaling, which means we get to stand on their infrastructure and spend our effort on the trust layer that is actually the differentiator. For a client who already lives on one of these platforms, the right answer is often to wrap our discipline around their runtime, not to rebuild it.
What we are not claiming
We are not claiming these platforms are weak. They are strong, well-built, and ahead of us on raw scale and on a few things worth learning from, first-class evaluation and monitoring among them. Naming where someone is ahead of you is more useful than pretending they are not.
We are not claiming the ninety-nine percent is easy because we said it out loud. It is the hard part precisely because it is unglamorous, invisible, and never the thing a buyer asks for first. The discipline is the work, whether you buy a platform or hire a firm.
We are not claiming you need us instead of a platform. Often you need a platform and someone who will hold the ninety-nine percent honestly on top of it. That second person is the job we are good at.
One-sentence takeaway
When two platform companies independently tell you the agent is one percent of the work, believe them, because the other ninety-nine, the security, the grounding, the cost control, the audit trail, and the human gate, is the entire product, and it is the part that was always going to be the work no matter who built the loop.
Talk to us
If you are evaluating an agent platform and trying to work out what it actually gives you versus what you still own, that is the conversation we are good at: the loop is handled, so the real questions are what is scoped, what is gated, what is grounded, what it costs, and who approves. We will tell you on the call where a platform already has you covered and where you still own the ninety-nine percent.
We do not take every engagement, and we will tell you whether we are the right partner.