Case studies / Agents
Agents
LLMs that use tools and work through multi-step tasks, when no narrower agent topic fits.
Approach
26 write-ups
Tag:
-
MongoDBNewAgents
The $5 Trillion Shift: Preparing Retail Backends for the Age of AI Agents
Problem Retailers struggled with fragmented databases for multiple AI agents, causing inconsistencies and slow responses.
Solution MongoDB offers a unified data platform that stores vectors, sessions, short‑ and long‑term memory, and audit logs in one place, enabling consistent, real‑time multi‑agent interactions.
Read on mongodb.comDetails -
KlaviyoNewAgentsSkills for Your Agents
Problem Teams needed a reliable way for AI agents to choose and execute domain‑specific workflows without constant manual oversight.
Solution Klaviyo created a skill system where each skill defines a workflow and resources; the model routes requests to the right skill, runs most steps automatically, and hands off only the final step to a human.
Read on klaviyo.techDetails -
PostmanNewAgentsMy vacation agent is struggling: how AI-ready are services?
Problem Comparing vacation offers across many travel sites is time‑consuming and confusing.
Solution Created a reusable agent skill that visits each site, extracts offer details, and scores them using a fixed formula, handling both API and browser access.
Read on blog.postman.comDetails -
NVIDIAAgents
Using AI Agents to Teach Robots Like Microduck to Walk in Simulation
Problem Robots struggled to learn walking and obstacle navigation without a coordinated planning and training system.
Solution Created a workflow where AI agents generate plans, build simulated environments, and train navigation policies using NVIDIA's RTX Spark, Nemotron, and Isaac Lab, then evaluate them in simulation.
Read on developer.nvidia.comDetails -
MicrosoftAgents
Flint-interactive: A language for creating interactive data visualizations
Problem Users wanted charts they could explore, not just static images.
Solution They created Flint‑interactive, a high‑level language that turns user or agent commands into semantic events, letting a compiler generate and update interactive visualizations automatically.
Read on microsoft.comDetails -
NVIDIAAgents
Design Proteins with RFD4-Proteina Using Reward-Guided Inference-Time Search
Problem Scientists need to explore billions of protein sequences to find therapeutic candidates, but exhaustive search is impossible.
Solution They created a model that generates protein sequences and structures together, then runs a reward‑guided search at inference time, keeping and expanding promising partial designs until the goal is met.
Faster and uses less memory than the previous version
Read on developer.nvidia.comDetails -
AppleAgents
Selection-Based Structured Reasoning: Toward Efficient Multimodal Search Agents
Problem Long, costly reasoning slowed multimodal search agents.
Solution They replaced freeform reasoning with a set of reusable reasoning options and let the model pick the best one, scoring all options in parallel.
Per‑turn latency dropped 90%, overall inference latency 28–54% faster
Read on machinelearning.apple.comDetails -
NVIDIAAgents
NVIDIA Nemotron 3.5 Super VL Powers Specialized Agents with Multimodal Reasoning
Problem Enterprise workflows need agents that can read documents, follow policies, and use multiple tools in one request.
Solution NVIDIA created Nemotron 3.5 Super VL, a 120B‑parameter MoE model that handles text, image, document, and video inputs, reasons, follows instructions, and calls tools, all on a single GPU.
Read on developer.nvidia.comDetails -
GoogleAgents
Putting Earth AI to work: Geospatial agents for our planet's biggest challenges
Problem Standard LLMs can’t handle huge, complex geospatial data and multi‑step spatial analysis.
Solution Created a Geospatial Reasoning agent that plans steps, picks the right Earth AI models, and calls tools like Maps, Earth Engine, and custom datasets to answer spatial questions.
Read on research.googleDetails -
DropboxAgents
Evolving our calendar assistant Reclaim to be AI-native without starting over
Problem Users wanted to schedule tasks by describing goals in their own words, not by picking preset settings.
Solution They built an AI agent that reads free‑text requests, checks the calendar context, and creates appropriate scheduling actions while reusing the existing scheduler.
Read on dropbox.techDetails -
SalesforceAgents
Engineering Multi-Agent AI Teams That Build and Test Themselves
Problem Creating and testing AI agents manually was slow and error‑prone.
Solution They built self‑modifying agents that can design, verify, test, and fix other agents, automating the whole development loop.
Read on engineering.salesforce.comDetails -
GoogleAgents
Automating coherent long-form video generation
Problem Long‑form video creation kept losing character look and story flow because separate tools failed and drifted.
Solution They built a multi‑agent system that plans and tracks visual continuity, coordinating diffusion models and other tools to keep scenes and characters consistent.
Substantial gains in multi‑shot narrative consistency and character persistence
Read on research.googleDetails -
DoctolibAgentsFrom hackathon to production in 3 weeks: inside Doctolib’s AI factory
Problem Clinicians had to rewrite AI‑generated notes because the style didn’t match their personal preferences.
Solution Doctolib built a Consultation Template Manager agent that lets doctors define note templates, which the ambient scribe follows to produce correctly formatted notes instantly.
Read on medium.comDetails -
GoogleAgents
The future of practice: Enabling teachers to create learning interactives with generative UI
Problem Teachers struggled to create engaging, interactive simulations for every lesson.
Solution They built a generative UI that lets teachers design guided, interactive simulations on demand, using learning‑design guardrails to keep content accurate and aligned with curriculum.
Read on research.googleDetails -
NVIDIAAgents
How to Use AI Agents to Prepare 3D Scenes for Simulation
Problem Robotics engineers spend lots of time making 3D scenes simulation‑ready, slowing development.
Solution They created a multi‑agent system where a central LLM coordinates sub‑agents that use Omniverse tools to inspect, add physics data, render previews, and output a SimReady OpenUSD scene.
Read on developer.nvidia.comDetails -
Hugging FaceAgents
Your Agent Aced the Task. Will It Do It Again?
Problem Agents sometimes give different answers to the same request, causing unreliable behavior.
Solution They added a Consistency Analyzer and new consistency guidelines to their ALTK‑Evolve system, which turn past agent runs into rules that are applied at inference to keep the agent on the same path each time.
Read on huggingface.coDetails -
GoogleAgents
Planetary prediction engine: Automating global models via Earth AI
Problem Manual data curation for geospatial models took weeks, slowing critical responses.
Solution They built an autonomous system that, from a natural‑language query, automatically finds relevant geospatial data, engineers features, trains and evaluates a model, and produces a report.
Reduced model build time from weeks to minutes
Read on research.googleDetails -
GoogleAgents
An AI tool for prioritizing candidate biomarkers from wearable sensor data
Problem Users needed trustworthy biomarkers from wearable data, but existing tools missed statistical rigor.
Solution They built a multi‑agent system that iteratively generates hypotheses, runs statistical tests, trains models, and checks literature to rank biomarkers, all under human supervision.
Recovered known signals, found new convergent biomarkers, and boosted prediction accuracy
Read on research.googleDetails -
GrabAgents
Building Jarvis Pro: Route first, answer later
Problem Account managers needed reliable AI advice for merchant meetings without giving wrong recommendations.
Solution Created Jarvis Pro that first classifies each user request, then loads only the needed data and generates an answer, using a routing guide to enforce safe paths.
Safe routing 99.4% accuracy, quality score rose to 91.0
Read on engineering.grab.comDetails -
Hugging FaceAgents
Record, train, and deploy from one place with Strands Agents, LeRobot, and Hugging Face Storage Buckets
Problem Repeated data uploads and downloads slowed training and increased costs as robot recordings grew.
Solution Created a continuous loop where robot demos are stored in Hugging Face Storage Buckets, read directly for training, and deployed back, avoiding redundant transfers.
Read on huggingface.coDetails -
ShopifyAgents
Clustering billions of products for agentic commerce with Catalog API
Problem Shopify needed to match the same real‑world products across billions of merchant listings that use different structures.
Solution They built an LLM‑driven pipeline that creates embeddings for each product, clusters them with a precision‑first threshold, and assigns a universal product ID so AI shopping agents can search a unified catalog.
Read on shopify.engineeringDetails -
Hugging FaceAgents
Beyond LLMs: Why Scalable Enterprise AI Adoption Depends on Agent Logic
Problem Enterprises struggled to use AI agents in complex, long‑running workflows with many APIs and strict policies.
Solution They built agents that use built‑in logic like knowledge graphs and algorithms to steer the LLM, keeping context small and improving reliability and cost.
Read on huggingface.coDetails -
KlaviyoAgents
Rethinking Software Design Principles for Agent Product Development
Problem Marketers need to create large volumes of multi‑channel campaign content that manual teams cannot produce fast enough.
Solution Klaviyo built Composer, an AI‑driven agent that takes natural‑language prompts, runs parallel generator agents, and outputs schema‑validated, platform‑aware, data‑grounded content for email, SMS, push and WhatsApp.
Read on klaviyo.techDetails -
SlackAgents
Managing context in long-run agentic applications
Problem Multiple long‑running agents lost alignment and context over time, causing incoherent behavior.
Solution Implemented context‑management mechanisms that synchronize agents, preserve shared state, and guide reasoning across extended interactions.
Read on slack.engineeringDetails -
GoogleAgents
Improving the academic workflow: Introducing two AI agents for better figures and peer review
Problem Researchers spend too much time making complex figures and reviewers are overwhelmed by paper volume.
Solution Created two AI agents: PaperVizAgent turns manuscript text into publication‑ready diagrams, and ScholarPeer writes detailed, literature‑grounded reviews, using a team of specialized sub‑agents.
Figures and reviews outperform existing automated baselines
Read on research.googleDetails -
GoogleAgents
How AI tools can redefine universal design to increase accessibility
Problem User interfaces were static and hard for people with disabilities to use.
Solution Created Natively Adaptive Interfaces that use multimodal AI agents to reshape UI elements in real time based on each user’s needs.
Read on research.googleDetails
No write-ups match these filters.