Prototyping is easy. Production is the bottleneck.

Model quality has collapsed the cost of generating code. Anyone can prompt Claude to scaffold a working agent in minutes. The same AI, however, will happily design a system that costs $5,000/month to operate when a $500/month architecture would do. The imbalance is no longer about building—it is about operating what gets built.

That shift changes the economics of internal tools. Custom software used to require a large engineering investment and long-term operational commitment before it paid off. For most companies, buying off-the-shelf was the rational default. Now, teams of any size can spin up specialized agents quickly, and the customization payoff is immediate. OpenAI runs an internal data agent for analytics, Vercel's lead qualification agent lets one SDR handle the workload of ten, and Stripe built a customer-facing financial impact calculator on a single flight.

The old build-versus-buy question is obsolete. The decision is now build and run. That requires treating internal AI workloads with the same rigor as production customer infrastructure—security, observability, reliability, and cost control. Those skills have not been democratized. Vibe coding has produced one of the largest shadow IT problems in history precisely because building is easy while production operations remain hard.

For agents specifically, the platform underneath determines whether a clever prototype becomes a dependable business tool.

How an internal data agent runs in production

Vercel's own analytics agent, d0, illustrates the gap between an agent idea and a reliable product. Text-to-SQL is not new. What made d0 work was the orchestration layer under it: using Vercel's platform primitives, one person built the agent in a few weeks on 20% of their time, skipping the months of scaffolding and security work that a custom stack would have required.

d0 answers natural language questions from Slack, drawing on a semantic layer of YAML configs that describe Vercel's warehouse, metrics, and operations. The request flow looks like this:

  • The user message arrives in Slack. d0 checks the user's permissions to determine appropriate data access, then starts the workflow.
  • The agent navigates a semantic layer. This file system of configuration describes the data warehouse and business definitions, so the agent does not guess at schema.
  • Model calls go through the AI SDK. Streaming, tool use, and structured outputs come from the standard abstractions available to any Vercel developer—no custom LLM plumbing.
  • Workflow steps are durable. On a Snowflake timeout or model failure, Vercel Workflows retries and restores state automatically instead of aborting the whole run.
  • Actions execute in isolated sandboxes. File exploration, SQL generation, and arbitrary Python for analysis run in secure Linux VMs. A runaway or compromised operation cannot touch core systems.
  • Routing balances cost and accuracy. AI Gateway sends simple queries to fast, cheap models and complex analysis to Claude Opus, all in one codebase.
  • The answer comes back to Slack as formatted results with charts or spreadsheet links via the AI SDK Chatbot primitive.

d0 has turned data access that once required a professional analyst into a self-service capability for engineers, marketers, and executives.

What an agent platform must provide

Vercel positions its orchestration stack as the runtime for both internal and customer-facing agents. The same primitives that run d0 and the lead qualification agent also power v0, the code generation product, and Vercel Agent, which reviews pull requests and analyzes incidents. The components address the specific failure modes of agent workloads.

Isolated execution

Agents generate and run untested code. They are also vulnerable to prompt injection. Sandboxes confine that risk inside isolated Linux VMs. When an agent needs filesystem access for discovery, sandboxes can mount VMs dynamically with controlled access to specific resources, protecting core systems from both accidents and attacks.

Elastic compute for unpredictable work

Agents do not have steady compute profiles. Text processing is light, but files, images, and video scale quickly, and costs follow. Fluid compute scales with demand and bills only for compute time actually used, keeping expenses predictable as workloads fluctuate.

Model abstraction and failover

No single model is optimal for every agent step. AI Gateway exposes hundreds of models through one interface with budget controls, usage monitoring, and load balancing. Simple requests route to fast models, harder analysis routes to stronger ones, and if a primary provider hits rate limits or an outage, traffic fails over to backups automatically. That also avoids locking an agent to one vendor's roadmap.

Durable orchestration

Multi-step agents fail. Providers time out, services hiccup, steps error. Workflows provides retry logic and state recovery at each step so an interruption does not restart the operation or demand manual intervention. For business-critical processes, that durability separates a toy from a tool.

Agent-level observability

Standard system metrics do not explain why an agent made a decision. Observability surfaces prompts, model responses, and decision paths. That visibility lets a team trace unexpected behavior, token overruns, or poor responses back to a specific model call or data source.

The agents Vercel runs internally—lead qualification, analytics, customer support handling 87% of initial questions, abuse detection, and content drafting from Slack threads—all rely on these same foundations. So do the customer-facing products. As more enterprises build their own versions of these tools, the differentiator will not be the agent code. It will be the infrastructure that keeps those agents secure, reliable, and affordable at scale. Companies that invest in that internal AI stack now will compound the advantage as their agent portfolio grows.