LLMOps Database Adds 419 Production Case Studies, Total Hits 1,182

What happened

ZenML's LLMOps Database is a curated archive of how companies actually run large language models in production. Its latest batch added 419 entries, bringing the total to 1,182 case studies.1 The update was published on December 15, 2025. The project started 17 months earlier with about 300 entries. ZenML says it has read and analyzed 7.3 million tokens of source material and generated 4 million tokens of summaries along the way.1 Days later, curator Alex Strick van Linschoten published a companion analysis, written once the archive passed 1,200 entries, that tries to pull lessons out of the pile.313 By spring 2026 he was citing more than 1,400 production deployments in podcast interviews, so the collection is still growing quickly.15

The new entries cover a wide range of industries. They include multi-agent systems in manufacturing, HIPAA-compliant retrieval-augmented generation (RAG) in healthcare, and very large consumer assistants.1 Even so, the pattern across the batch is clear. Teams that ship agents with measurable results are winning on infrastructure discipline, not on access to the newest model.

The numbers companies are reporting

The batch contains plenty of concrete results, and they show that agentic systems now carry real traffic. Alan, a health insurer, says agents automate 30–35% of customer service for its 1 million members at quality comparable to humans. It also says 60% of reimbursements are processed in under five minutes. Getting there took a multi-model setup, prompt caching, speculative execution and refactored APIs to stay under a two-second latency target.1 An Infosys Topaz deployment for an energy supplier reportedly handles 70% of help-desk calls and cut average handling time by 60%.1 Prime Video's artwork moderation system, built on multimodal models and Strands agents, is said to have cut manual review by 88%.1

Some of the strongest results come from plain cost engineering rather than agent design. Care Access caches large medical records as a static prompt prefix on Amazon Bedrock, so only the changing questions get processed. It reports 86% lower inference costs and 66% faster processing.1 Cursor used online reinforcement learning on its Tab completion system, which handles more than 400 million requests a day, with user accepts and rejects as the reward signal. The result was 21% fewer suggestions and a 28% higher accept rate.1

These figures come with a clear limit. They are self-reported, often from vendor-sponsored conference talks, and the database summarizes them rather than checking them. The curators say as much in places. One entry on Factory.ai notes that the lack of specific metrics makes the real production impact hard to judge.16

Infrastructure is where the work happens

The main argument of the companion analysis is simple. The bottleneck is engineering, not model intelligence. The skills in demand are distributed systems, networking, infrastructure management and building reliable systems out of unreliable parts.13 Read on their own, the case studies back this up.

Take durable execution. The analysis points to teams using Temporal so that agent workflows survive backend failures and resume where they stopped. It describes this durability as the line between agents that work in production and agents that only work in demos.13 Vercel's AWS re:Invent session makes the same case from the vendor side. It cites a figure that 46% of AI proofs of concept are cancelled before production, and names model switching, evaluation and observability as the areas teams most often neglect.17 Union's write-up of Dragonfly describes a tiered agent architecture that runs 250,000+ software products through 200+ steps and 100+ LLM calls each. It reports more than 2,000 concurrent runs and a 50% cut in failure recovery time.16

Lower-level integration problems show up too. BrainGrid ran a multi-tenant Model Context Protocol (MCP) server on serverless infrastructure. Stateless instances kept forcing users to log in again and slowed token validation. The fix was a Redis session store with AES-256-GCM encryption and cached JWT validation.1 Amazon's Rufus team found that one machine could not hold the inference workload. It built a multi-node setup across tens of thousands of Trainium chips, using vLLM, a leader/follower design and network-aware node placement.1 These are distributed-systems problems more than AI problems, which is the analysis's point.

The analysis also suggests MCP is settling into a modest role, one year in. It now works as a standard way to connect agents to existing systems, not as a solution in its own right. The analysis compares it to USB-C: useful precisely because it is only a connector.13

Guardrails move into code, and context beats prompts

The clearest architectural lesson is that safety controls are moving out of prompts and into code.3 Oso uses "session tainting": once an agent has touched both untrusted content and sensitive data, it is blocked from using communication tools, whatever the model tries to do. Wakam runs two separate permission layers, one for what humans can see and one for what agents can access. Komodo Health removed all authorization logic from its model entirely.3 The most mature teams also add circuit breakers, hard limits that stop an agent automatically when certain thresholds are crossed. The reasoning is that models will sometimes misbehave even after thorough testing.13

The second lesson is that context engineering, meaning what information reaches the model and how, now matters more than prompt wording.2 Manus runs agents inside full virtual-machine sandboxes. For many tasks it skips the vector database and lets agents read the Linux file system with grep, cat and ls, writing bulky tool output to files instead of the context window.13 Elastic's presenter estimated that context engineering is about 80% agentic search. He found that the quality of a tool's written description is the biggest factor in whether an agent picks the right tool, and that most developers spend little effort on it. Cubic saw the opposite of what more capability should bring. Giving its code-review agent more tools made it worse: false positives piled up until developers stopped trusting it.13

How the picture has changed

Comparing this batch with ZenML's earlier write-ups shows how far the field has moved. In late 2024, the curators concluded that agents were potentially very powerful but "deeply unready for primetime".4 At the time, the success stories mostly involved pulling back from autonomy. Lindy.ai, for example, became much more reliable after switching from open-ended prompts to structured workflows.7 The December 2025 analysis is far more confident about production use. Ramp's policy agent, for instance, now handles over 65% of expense approvals on its own, while explaining its reasoning and handing uncertain cases to humans.13

The two periods don't contradict each other. Agents became usable mainly because teams wrapped them in constraints: human escalation, limited tool sets, guardrails enforced in code and durable workflows. Alan's move from fixed workflows to more flexible playbooks fits this pattern, but it still runs on a heavily engineered base.1 The analysis itself says agent harnesses, memory systems and reinforcement learning remain unsettled. They show promise but have not become consensus patterns, and are hard to reproduce outside the major labs.13

What comes next looks less stable than the case studies suggest. In an April 2026 interview, Strick van Linschoten said the most successful teams rebuild their agent systems every few weeks as models improve, and warned that over-engineering now creates technical debt.15 He also flagged "silent failures" as the biggest production risk: agents report success while running up costly mistakes. His example was a $50,000 infinite loop.15

Who is telling the story

Any reading of this archive has to account for who built it. ZenML sells MLOps orchestration software. Its own material argues that an MLOps backend, with versioned prompts, lineage and evaluation baselines, is the natural home for agents. A database whose main finding is that infrastructure discipline wins is convenient for a company selling infrastructure. The sample also leans heavily on AWS conference sessions, which affects which tools appear most often.117

That said, the conclusion holds up beyond any one vendor. Google's Vertex AI team reached a similar view independently. Patrick Marlow, a staff engineer there, predicted that models will become commodities and that the ecosystem around them will set deployments apart. He called evaluations the one thing teams should do even if they do nothing else.

For teams building agentic systems, the practical takeaway is that model choice matters less than the surrounding work: model-switching layers, stage-by-stage evaluations, guardrails enforced in code, durable state and cost controls such as caching. The case studies don't prove these practices guarantee success, since the evidence is self-reported. But across more than a thousand deployments, they are the common thread among teams that report results in production rather than just good demos.