AI 服务暂时不可用，以下为来源正文，待恢复后补全翻译。
Case Studies How Numan built an eval-first AI health coach with LangGraph and LangSmith October 19, 2026 min Go back to blog Create agents Create agents Numan is a UK digital health company helping patients manage long-term conditions across weight, hormones, and sexual health. Many of our patients are on a weight-loss programme and use the app daily to track progress. When building Nu, our AI health coach, one of our design challenges was delivering high-quality health coaching without crossing the line into medical diagnosis, prescriptions, or making clinical decisions—all strict boundaries for agents in our highly regulated medical space. While we initially focused on the agent and automation around it, what turned out to matter most was the quality of our evaluation criteria. This is the story of how we developed and improved our evaluation criteria to deliver high-quality user experiences while steering clear of areas subject to regulation. Where standard eval approaches fell short for Nu We started where most teams start, with a dataset of inputs and expected outputs, and an LLM judge scoring each reply. While that playbook works well for a Q&A agent, coaching is a conversation, not an answer. Every approach that fixed Nu's reply in advance failed in two ways: 1. Single-turn evals on static fixtures (a fixed input, a fixed expected output, and a judge scoring one reply on its own). Behavioural coaching rarely has one right answer to a prompt and whether a reply was safe depends on the conversation around it. Nu needs to understand what is getting in the patient’s way across a conversation before it advises and hold the right boundary when a patient pushes. Averaging one score across the whole dataset hid where Nu failed. A few bad replies barely move the number, so the conversation looks compliant when it isn’t. The fixtures also caused a second problem. In one run, a cluster of low scores turned out to be the eval’s fault, not Nu’s. The rubric expected empathy, exploration and advice all within a two-turn exchange, whereas a human coach would pace these across a conversation. Nu was coaching properly, but the narrow, fixed eval signaled failures incorrectly. 2. Scripted multi-turn conversations. The obvious fix for single-turn fixtures is to write out a patient’s expected series of replies in advance. With this "script," several turns of a scenario, fed to Nu one at a time, allow the judge to score the full transcript at the end. While this gives the judge more room than a single reply, the eval fell apart the same way. When the patient’s second turn is written in advance, the test ignores what Nu actually said in the first turn. The moment Nu says something the script did not expect, every later turn is replying to a conversation that never happened. These 2 failure modes shaped everything we built thereafter. The eval framework we run today Nu is a LangGraph graph. A supervisor model orchestrates handoffs to domain-specialist sub-agents, each focused on one coaching area with its own prompt and tools. That composition model makes a new specialist cheap to add, and more importantly, it keeps the evaluation surface uniform across all of them. Some of what Nu has to get right is decidable from a single reply. For example, whether the supervisor routed to the right specialist, whether a medication question got a safe answer, or whether the safety guard caught what it should. For those we run themed datasets spanning coaching quality, safety, medication queries, routing, and summarisation. Unlike our original single-turn evals, we changed two things: 1) Every dataset is scored per-category by an LLM judge rather than one average, so a cluster of failures in one category stays visible even when the rest look healthy. And 2) each rubric asks only what a single reply can show, so the judge no longer penalises Nu for pacing a conversation the way a coach would. Everything that a fixture cannot cover (i.e. what requires the shape of a whole conversation) gets its own layer, described below. Because Nu is a graph, an eval can target the whole system or a single sub-agent, and a new specialist inherits the eval surface. "The thing we ship is the thing we test" The easy way to build an eval is a separate harness that copies the agent’s prompt and calls the model directly...which works until the harness and production drift. We have no separate harness and operate under the principle the thing we ship is the thing we test. Every eval runs through the same LangGraph graph a patient talks to, so what we score is what a patient meets. As we scaled, we started asking whether one provider was right for every agent, since the supervisor and the domain specialists do different cognitive work. Nu runs on two frontier providers today, chosen per-agent: one for the supervisor, another for the specialists. Thanks to the portability of our eval suite built as open, themed datasets we can point at any configuration, we ran the same datasets across providers and read the results side by side. A migration between frontier providers, normally weeks of manual testing, took only a day. Behavioural tests to keep the regulatory boundary Under MHRA guidance, a software-based product with a medical purpose (giving clinical advice, diagnosing, or recommending treatment) becomes a medical device. Nu was designed to be a wellness product for lifestyle coaching and education, not clinical care. Keeping Nu inside that scope is something we check on every change. Safety probes run as per-category metrics and gate releases: a regression there blocks the change from shipping automatically. We tune toward safe-completion, a redirect that holds the line while still sounding like a coach, so a patient who asks something out of scope gets a warm, useful answer rather than an abrupt refusal. Beyond testing, guard nodes run on every conversation inside the graph. A safety guard reviews each outgoing response against a fixed set of clinical safety rules before the patient sees it. An unsafe draft is sent back to the supervisor to regenerate. If it keeps failing, the message is replaced with a safe fallback rather than delivering the unsafe response. And because the guards sit in the graph, they sit inside the same eval surface—where one of our themed datasets exists specifically to test that the safety guard catches what it should. In addition, health coaches review a sample of real conversations every day and flag anything that looks like a safety miss, so human subject matter experts calibrate what counts as safe. Simulating conversations instead of scripting them Our newest layer, in prototype today, simulates the conversation instead of scripting it. We hand the eval a patient persona, a scenario, the key problems Nu should draw out, and a turn limit. A cheaper model plays the patient, holding a multi-turn conversation with the real agent and reacting to whatever Nu says. Then a judge scores the whole transcript per-category. Because the patient is generated rather than fixed, we can probe adversarial cases (phrasings built to slip a clinical request past the coach, manipulation disguised as an ordinary question) and catch failures a fixture never encounters. For example, Nu rushing to an answer instead of investigating the way a coach should. This kind of prompt hardening via simulated conversations reduced serious scope violations in jailbreak scenarios against the regulatory boundary. We ran parallel medication-query evals as a regression gate so prompt improvements could not break behaviour we already trusted. Using the simulated conversation layer for continuous improvement without introducing regressions is less an engineering problem than a rubric problem. When we first drafted criteria for this layer, two themes we thought were compatible turned out to contradict each other on missed-dose handling: one said decline and signpost to a clinician, the other said safe-completion with general facts from the patient information leaflet. One prompt cannot satisfy both. This example of a prompt conflict taught us that the scarce, valuable asset is not the prompt, which a capable model drafts well from limited context; it is the eval criteria. And on a regulated boundary, those criteria have to be authored and signed off by a clinician who also understands the regulatory line, not an engineer. The results from Nu's simulated conversations are judged against engineering-drafted criteria, which is why this layer does not gate releases or inform any safety claim until a clinician has signed off the rubric (which is what keeps the product on the wellness side of the line). The stack: LangGraph and LangSmith Real and simulated conversations run through the same LangGraph agent; LangSmith measures both. A clinician-owned rubric drives the judge, and a boundary regression blocks the release. Dashed = prototype or direction. LangGraph and LangSmith fit together as one system to make up our agent development stack. LangGraph is the agent. A supervisor triggers domain specialists, each with its own prompt and tools. Evals run through that same graph. Both fixture and simulated conversation eval layers hit the real agent and can target the whole system or a single sub-agent. LangSmith is where everything is measured. Every conversation, simulated or real, lands as a trace. Reviewers annotate those traces, kicking off the continuous quality improvement loop. Datasets and experiments live there too. For example, the side-by-side provider comparison behind our one-day migration was read straight out of LangSmith. The integration between LangGraph and LangSmith keeps evaluation cheap to run. There is no glue code stitching three disconnected tools together. The thing we ship and the thing we test are the same graph, watched by the same platform. Where we are heading: scaling the criteria, not just the agent Coaching quality hinges on 1) authoring the criteria that decide whether a conversation was good, and 2) widening the coverage they are tested against. So we're focused on testing more adversarial personas and scenarios, adding more themes as we build more specialists, and expanding the clinician-in-the-loop rubric wherever it touches safety or scope. Reviewers keep reading real conversations in LangSmith, and an engineer still raises every prompt change. Everything described so far is offline, proving a change holds the regulatory line before it ships. The next step is online evaluators that run the same rubrics continuously against real conversations as they land in LangSmith, scoring Nu's side of the conversation. Instead of the daily sample coaches review today, human reviewers will start from the conversations an evaluator has already flagged—to watch the boundary hold in production and sharpen the rubric upon false alarms. We're also seeing encouraging results from early experiments to automate the workflows that follow a trustworthy rubric—including triaging reviewed conversations into a ranked backlog and tuning a prompt against a fixed benchmark. Recently, an offline prompt-optimisation loop pointed at a single behavioural benchmark meaningfully improved the score with no hand-tuning...in favor of our hypothesis that if you get the criteria right, the rest becomes automatable. Building an AI coach near a regulatory line If you are building a conversational agent anywhere near a regulatory line, here are our recommendations: A fixed dataset cannot cover a whole conversation. Keep fixtures for what a single reply can show, simulate conversations for the rest, and score per-category rather than chasing an aggregate. Make the boundary a probe. Run it on every change, let a regression block the release, and aim the same rubric at production. On a regulated boundary, the rubric belongs to a clinician. Engineering-drafted criteria are enough [truncated for AI cost control]