Production agent architectures are converging around federated systems, specialized classification models, and cost-aware routing layers. This week's reads show how teams are shipping customer-facing agents while managing the token spend and architectural complexity that comes with multi-model deployments.
How Tailscale built a customer-facing model router on AI Gateway
Tailscale's Aperture demonstrates the build-versus-buy inflection point for model routing infrastructure. Instead of building their own gateway layer, they leveraged Vercel's AI Gateway to access hundreds of models through a single API with zero markup and no data retention, while Sandbox handles secure execution inside customer tailnets. The migration story is the real signal here: moving their entire company's AI usage took seconds with zero disruption, which matters when you're shipping a product customers will integrate into their own networks.
[AINews] Reality Checks on AI News
Steve Yegge's Gas Town shutdown and Databricks' 60% cost increase after rolling out GPT-6 Astra tell the same story from opposite ends: performance improvements don't automatically translate to productivity or savings. Databricks saw Astra outperform on complex tasks but still watched overall spend climb as 3,500 engineers adopted it. Meanwhile, Yegge admitted he never successfully built anything despite thousands spent on coding agents. The gap between capability and outcome is where observability and budget controls actually matter.
Building a Harness with Jev
Jev is a System One model optimized for classification rather than generation, offering 200x faster inference and 400x lower cost than traditional LLMs for decision-making tasks. The architectural shift is interesting: instead of generating text and parsing it, Jev evaluates multiple questions in parallel and returns typed probabilities for routing, safety checks, and mode selection within agent loops.
const classifier = new TypeSafeClassifier({
model: "jev-1",
questions: {
safety_risk: { type: "yes_no", question: "Does this request pose a safety risk?" },
route: { type: "choice", question: "Which model should handle this?", choices: ["fast", "capable", "vision"] }
}
});
const result = await classifier.invoke({ input: userRequest });
// Returns: { safety_risk: 0.02, route: { fast: 0.85, capable: 0.12, vision: 0.03 } }This pattern lets you build cost-aware routing without burning tokens on generation for every decision.
Scaling Agents in Healthcare & Life Sciences
Healthcare organizations require tracing, evaluation, and spend visibility before granting agents autonomy (76% in this survey), which makes them a leading indicator for enterprise production patterns. Abridge processes 100M+ conversations annually with automated evaluation pipelines and production monitoring, while 49% of healthcare orgs are building centralized agent platforms to consolidate what started as fragmented experiments. The compliance requirements force architectural discipline that other industries will eventually adopt.
How Included Health Built Federated Agents for Healthcare Navigation
Included Health's Dot uses a supergraph architecture where different product teams own domain-specific workflows that share context through Deep Agents' global prompts and filesystem. The production metrics (75% engagement lift, 95%+ clinical routing accuracy, 99% high-risk detection) matter less than the architectural pattern: federated ownership with centralized orchestration via LangGraph's durable execution. This lets teams ship independently while maintaining consistency, which is the only way multi-agent systems scale across organizations.
The pattern across these posts is clear: production agents require routing layers, classification models for fast decisions, and federated architectures that let teams own domains without fragmenting the user experience. Cost control and observability aren't optional extras, they're what separates pilots from production.