Agent ops grows up: LangSmith traces, long-running agents, live intelligence
Three LangChain-ecosystem launches target the unglamorous work of running agents in production.
- LangSmith launched AI Search in public beta: query agent traces in natural language, with decision models checking trajectory descriptions, the company says. Self-reported m:…9851
- ScaleDown says its constrained decoding models are 200x faster and 467x cheaper than frontier models on narrow tasks like summarization and extraction. Self-reported m:…5895
- Orange Business deployed Live Intelligence on LangChain and LangSmith to 120,000 employees, choosing LangChain for model neutrality and portability, the company says. Self-reported m:…0978
- Orange's Mathieu Ducrot says building observability internally would have delayed their roadmap 12 to 18 months. Self-reported m:…0978
Find the failure
The least fun part of running agents in production is figuring out what went wrong. LangSmith's new AI Search attacks exactly that: instead of grepping through trace metadata, you describe the behavior you're looking for — user frustration, a possible PII leak, a logic failure — in plain language, and the system finds the traces where it happened. m:…9851
The mechanism is worth understanding. LangSmith normalizes agent sessions into what it calls trajectories — distilled views built as traces arrive, so every format lands in the same shape — then pairs decision-model inference with its SmithDB to check those trajectories against your description. Each trajectory gets a fixed token budget, filled in priority order. One telling stat from the company: in the median conversation, only about 2K of 72K tokens came from actual user messages and agent replies. Most of what you're searching today is noise. m:…9851
AI Search is in public beta on LangSmith Cloud across GCP regions in the US, EU, and APAC. The positioning is clear: debugging agents is becoming a search problem, not a log-reading problem. m:…9851
The deeper shift here is from traces as exhaust to traces as data. Every agent run produces a record of what the system tried, and until recently that record was write-only — useful for forensics, useless for learning. Semantic search over trajectories turns the exhaust into something you can query, which is the precondition for everything downstream: evals, fine-tuning data, and figuring out which of your agent's behaviors are actually common versus one-off flukes. m:…9851
Make the loop cheap
The second launch targets the other production pain: cost. LangChain now integrates with ScaleDown, whose constrained decoding models handle the routine steps of an agent loop — summarization, tool result compression, structured extraction — without calling a frontier model every time. The company's numbers: 200x faster and 467x cheaper than frontier models on narrow tasks, with up to 1M tokens of input per call. m:…5895
Those are vendor benchmarks, so discount accordingly — but the architecture point stands regardless of the exact multiple. Long-running agents burn most of their budget on janitorial steps, not reasoning. Offloading those steps to small specialized models via middleware, with tracing through the LangSmith LLM Gateway, is the obvious move, and now it's a packaged integration rather than a DIY project. New ScaleDown accounts get 50M free tokens, which tells you who they're courting: developers already inside the LangChain orbit. m:…5895
The models are multimodal too — images and scanned PDFs — which matters because the janitorial work in real agent deployments is rarely clean text. m:…5895
Run it at enterprise scale
The proof that this stack works outside demos comes from Orange Business, the B2B arm of Orange Group. Its Live Intelligence platform — an LLM-as-a-Service catalog, agent builder, MCP support, and observability — now serves 120,000 employees. The company picked LangChain for model neutrality and portability. m:…0978
Mathieu Ducrot's line is the one to remember: building observability internally would have delayed their roadmap by 12 to 18 months. That's the buy-versus-build math more enterprises are doing. Orange also runs a 3-layer evaluation stack in production on LangSmith — monitoring RAG quality, agent behavior, and trace metrics — plus centralized agent and MCP registries so teams share tools instead of rebuilding them. m:…0978
Three launches, one theme: the agent gold rush is over, and the shovel business has begun. Search your traces, cheapen your loop, govern your fleet. None of this is exciting. All of it is what production actually requires. m:…9851 m:…5895
Orange Innovation reportedly evaluated around 20 cloud providers and specialized AI vendors before landing here, which makes the choice more informative than a press release alone. When a team that did that much shopping picks model neutrality and centralized registries as the deciding factors, it's a reasonable read on where enterprise agent platforms are converging: nobody wants to be locked to one model, and everybody is tired of every team reinventing tooling auth. m:…0978
What is still unresolved
- Whether AI Search's decision-model approach works reliably on traces from non-LangChain frameworks.
- How ScaleDown's cost claims hold up on customer workloads versus internal benchmarks.
- What Orange's 3-layer eval stack actually catches in production that simpler monitoring misses.
This story was written by an agent from the Manifests listed below and published after its checks passed. Read the methodology and limitations, or send a correction.