Production AI is no longer theoretical. This week's releases from LangChain, Vercel, and Shopify all center on the same challenge: making agentic systems observable, durable, and developer-friendly enough to ship without fear.
How Pendo used LangSmith to trace Novus from user behavior to code fixes
Pendo built Novus, an agent that watches user sessions, detects usability friction, and generates code fixes with a 90%+ success rate. The real story is LangSmith's observability layer catching 60% of agent failures before users report them. Tracing revealed that the agent initially failed to combine analytics signals with code context, a prompt tuning win they only discovered because every decision was instrumented. This is what production agent workflows actually look like: tight feedback loops between telemetry and iteration.
// LangSmith tracing reveals which context was missing
const trace = await client.readRun(runId);
if (!trace.outputs.codeContext) {
// Agent saw analytics but not code structure
console.log("Missing code context in usability fix");
}How Candidly Built State-Aware Agent Harnesses with LangSmith
Candidly turned post-mortem evaluation into real-time steering. Their Input-Output Hidden Markov Model infers four engagement states (Engaged, Detailed, Guided, Disengaging) from conversation traces and dynamically adjusts response strategies mid-conversation. Disengaging conversations dropped from 23% to 11%. The architectural insight: separating user-side signals from agent-side levers means the agent can adapt without waiting for explicit feedback or manual retraining.
Vercel Services: Run full stack on Vercel
Vercel Services lets you co-locate Next.js frontends and FastAPI backends in one project, with atomic deployments and internal service bindings that skip the public internet. Services are declared in vercel.json, and communication happens via injected environment variables.
{
"services": [
{ "name": "web", "framework": "nextjs" },
{ "name": "api", "framework": "fastapi", "bind": ["web"] }
]
}This solves the coordination tax of multi-repo full-stack projects and makes preview environments actually useful. Unified compute, WebSockets, workflows, and isolated sandbox environments for agents all land in the same deployment surface.
AI SDK 7
AI SDK 7 ships WorkflowAgent for durable execution that survives restarts, standardized reasoning control across model providers, provider-agnostic file uploads to reduce bandwidth waste, and OpenTelemetry integration for multi-step agent observability. HarnessAgent allows you to wrap third-party agents like Claude Code and expose them through the same interface. The design goal is clear: make agentic workflows as predictable and debuggable as stateless API calls.
Vercel and Shopify are rebuilding Hydrogen
Hydrogen is being rebuilt as a runtime-agnostic, open-source architecture with three layers: core (centralized Shopify API logic), client (reusable cart and checkout components), and server (framework integrations for caching and performance). You can now use Hydrogen with Svelte, Nuxt, or Next.js instead of being locked into a single stack. This eliminates repetitive glue code and makes Shopify storefront development feel like choosing a UI library instead of committing to an entire platform.
The throughline: observability, durability, and portability are no longer nice-to-haves. If your agent doesn't trace, your full-stack setup requires four repos, or your framework locks you into one runtime, you're competing with tools that just solved all three. 🛠️