Frontier labs and hyperscalers promise world-shifting innovation. And most deliver it. But, as we’re seeing with the Anthropic policy flip-flop and the evolving Hugging Face and OpenAI security incident, they operate without stability.
That’s deeply concerning because technology organizations that build their entire AI operations and business on top of Anthropic, OpenAI and other paid models need to be able to depend on their reliability.
Meanwhile, open-source organizations like OpenClaw and DeepSeek offer cost-free models with similar quality. The difference in price is stark. And the gaps in utility, safety and accessibility that kept the enterprise away are closing fast.
This evolving dynamic leaves CTOs, CIOs and engineering leaders with a question: How can we keep reliability up and costs down when it’s impossible to predict whether hyperscalers will drop or raise prices of their next models?
The answer isn’t clear-cut — yet. But it’s never been clearer that engineering leaders need systems that allow their teams to quickly swap models and shift how AI agents work with people and access real data and tools. Building the right foundational layer keeps organizations nimble enough to evolve alongside the industry without cutting corners by chasing the latest trends.
Tokens cost more than time and money
Engineering leaders at Big Tech companies and within enterprises learned the hard way that building toward their organization’s long-term stability is a much better plan than chasing trends like “tokenmaxxing,” which results in unsustainable spend and team burnout.
While a fair amount of damage to company accounts and executive reputations has been done, the pendulum is already swinging back from tokenmaxxing to more sober approaches. At the same time, companies like Meta that publicly went all-in on team-wide AI use are shifting toward reinvesting in engineering team culture.
The goal: boosting morale while removing competition from token use.
Instead of jumping on the next hype train and creating the inevitable bottleneck, organizations should invest in modernizing their infrastructure to empower teams to sustainably iterate on and experiment with AI tools at scale.
The future of enterprise AI empowers people and agents to work seamlessly together. What this looks like:
- Accepting that agents have most of the same capabilities as people, with the added value of being able to test against real infrastructure with access to “real” data swiftly and at scale.
- Ensuring agents have the same guardrails as teams, including making sure credentials and permissions are only granted when needed; under the right circumstances and with full visibility into actions taken when things go wrong.
- Building systems that are able to swap in the latest AI models and frameworks to take advantage of new advancements without losing the custom work done in-house.
- Making sure their companies aren’t locked into a single provider long-term in order to reduce risk from outages, expensive contracts or dated products.
Models change. Update your architecture
Building resilience starts with accepting that models and how we use them will change. Engineering leaders need to embrace that it will sometimes make sense to go with the latest hyperscaler model. Other times, it will make sense to bring in open-source models with novel harnesses that run at no cost but change how people collaborate with them.
Meanwhile, agents shouldn’t be limited to toy problems or synthetic environments. They need the ability to test against real infrastructure, interact with realistic datasets, and participate meaningfully in real business workflows.
The winning approach: Level the playing field between agents and engineers.
Give agents access to the same environments people use and mandate that they operate under the same guardrails teams follow. Permissions should be granted only when necessary. Credentials should be tightly controlled. Every action should be observable and auditable. When something goes wrong, accountability should follow with clear visibility into what happened and why.
Hold both parties to the highest standards. Build resilience with your team.
There’s strength in flexibility
The days of custom workflows, automation and operational knowledge being trapped behind a single vendor relationship are over. We’re entering an AI agent-plus-engineer era that demands building systems and teams around flexibility, elasticity and adaptability.
In other words, it’s time to eliminate long-term lock-in for good.
Organizations that preserve the flexibility to adopt new models, integrate emerging tools, and respond to changing market conditions without rebuilding everything from scratch build resilience with every model release. It’s the way of the future. Engineering leaders should adopt this approach today.
Sumeet Vaidya is the CEO and co-founder of Crafting, which aims to bring enterprise quality infrastructure to autonomous agents and engineers. He was previously an early engineering leader at Meta, Uber and Discord.
Illustration: Dom Guzman
