Case studies / Agent platforms
Agent platforms
Running agents at scale: harnesses, sandboxes, memory, MCP, observability and training.
Approach
64 write-ups
Tag:
-
AWSNewAgent platformsBuild autonomous AI agents with Amazon Bedrock and Spring AI
Problem Java teams struggled to move AI agents from prototype to production due to missing scalability, governance, and security.
Solution They used Amazon Bedrock AgentCore and Spring AI to build, deploy, and run autonomous agents with tool calling, memory, and multi‑agent coordination on existing JVM pipelines.
Read on aws.amazon.comDetails -
MongoDBNewAgent platforms
Introducing the MongoDB Claude SDK Session Adapter: Durable Session Storage for Agents
Problem Agents lost conversation history when running on stateless, disposable compute instances.
Solution MongoDB created a Session Adapter that stores Claude agent transcripts in MongoDB, letting any instance resume the exact session state across restarts and devices.
Read on mongodb.comDetails -
SnowflakeNewAgent platformsSame Table, Different Revenue: Apache Ossie and Snowflake to the Rescue
Problem Analysts and dashboards get inconsistent revenue numbers because the meaning of the data isn’t clearly defined.
Solution Snowflake created semantic views that embed business rules (like only completed orders count) into the data model, so queries, dashboards, and AI agents all use the same definition.
Read on medium.comDetails -
Hugging FaceNewAgent platforms
BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model
Problem Cloud LLMs are expensive, unpredictable, and expose user data, while many tasks need only modest intelligence.
Solution They built BotPsy, a 0.8 B‑parameter on‑device model that can call tools and act as a personal assistant for many everyday tasks.
Outperforms all models ≤1.2 B on tool‑use benchmarks
Read on huggingface.coDetails -
DXC TechnologyNewAgent platforms
How DXC Technology built a multi-model AI substrate on Amazon Bedrock
Problem Regulated enterprises needed AI agents that could act safely on critical systems without causing unintended damage.
Solution DXC built a governed model substrate on Amazon Bedrock that provides multi‑model access, built‑in guardrails, VPC‑only traffic, and serverless scaling, then added its own orchestration layer for enterprise IT operations.
Resolution time reduced 95%
Read on aws.amazon.comDetails -
AWSNewAgent platforms
Automating AWS Agent Registry updates with push-based architecture
Problem Registry metadata became stale, causing developers to lose trust in the catalog.
Solution Created an event‑driven pipeline using CloudTrail, EventBridge and Lambda that automatically syncs MCP server changes to the Agent Registry across accounts.
Read on aws.amazon.comDetails -
NVIDIANewAgent platforms
Agent Substrate: Advancing Kubernetes for Agents at Scale
Problem AI agents spend most time idle or waiting, causing wasted resources and slow response due to Kubernetes pod startup.
Solution Created Agent Substrate, an open‑source runtime that pools a few Kubernetes pods, checkpoints idle agent state, and restores it instantly, enabling millions of agents to run with high density.
Read on developer.nvidia.comDetails -
VercelNewAgent platforms
AI Gateway adds DeepSeek Search and Fetch tools
Problem Developers needed a simple way to let LLMs fetch web pages and search results during tool‑calling tasks.
Solution Added DeepSeek Search and Fetch tools to AI Gateway, letting any model that supports tool calling retrieve page titles, URLs, summaries, and markdown content via a single API key.
Read on vercel.comDetails -
KlaviyoNewAgent platformsInside the Harness: Deterministic Systems
Problem Managing massive context and ensuring AI agents follow strict guardrails was hard.
Solution They created a cascading disclosure system, routing rules, token caps, and pre‑tool hooks to keep the analytics agent deterministic and efficient.
Read on klaviyo.techDetails -
DropboxAgent platforms
How two engineer side projects became the agent taking action inside Dropbox
Problem Users wanted AI to edit their Dropbox files but needed safety and control over changes.
Solution Created a sandboxed agent that connects to Dropbox via rclone, uses permission checks and version history to safely modify files and allow rollbacks.
Read on dropbox.techDetails -
NeonNewAgent platformsInside a Kylon room, where people and agents build apps on Neon
Problem Teams needed a way for people and AI agents to work together on apps safely and instantly.
Solution Neon powers Kylon rooms where agents and users share a live view, using cheap database branches for drafts, reviews, and fast publishing without console access.
Read on neon.comDetails -
NVIDIANewAgent platforms
Use NVIDIA NeMo to Turn Agent Failures into Tests and Harness Fixes
Problem Agents sometimes pick the wrong tool or miss context, causing task failures.
Solution They built a loop using NeMo Compass to turn trace logs into prioritized test cases, wrote evaluations with NeMo Eval Author, and refined prompts, routing and memory to fix the agent.
Read on developer.nvidia.comDetails -
Mozilla.aiNewAgent platformsOtari by Mozilla AI: your operating system for AI governance
Problem Teams struggle to control AI usage, costs, and safety while being locked into specific providers.
Solution Mozilla AI built Otari, a gateway that enforces model, tool, and budget rules per workspace, logs usage, and lets apps switch providers via config.
Read on blog.mozilla.aiDetails -
ElasticNewAgent platformsHow our agent skills research, build, and test a complete Elastic integration
Problem Creating end‑to‑end Elastic integrations was manual and slow, requiring custom code for each data source.
Solution Elastic built an orchestration agent that coordinates specialist sub‑agents to research, generate CEL programs, set up ingest pipelines, run tests and review, automatically producing a validated integration.
Read on elastic.coDetails -
VercelNewAgent platforms
Vercel MCP now renders deployments as an interactive card
Problem Developers wanted to see deployment status and logs directly inside chat without switching tools.
Solution Created an interactive card for ChatGPT that shows live deployment info and lets the model fetch logs via an open_deployments tool.
Read on vercel.comDetails -
VercelNewAgent platforms
Vercel MCP now supports MCP events
Problem Developers had to poll Vercel for deployment updates, wasting time and resources.
Solution They created an event system that lets agents subscribe to deployment events via managed webhooks, handling filtering, expiration, and limits automatically.
Read on vercel.comDetails -
VercelNewAgent platforms
The fx adapter now runs the harness on the host
Problem Setting up sandboxed tool execution was complex and kept credentials insecure.
Solution Moved the harness to run in the host process with libfx, isolating filesystem tools, adding tool filtering, mid‑turn steering, structured output, and session resume.
Read on vercel.comDetails -
VercelNewAgent platforms
Run Paseo agents on Vercel Sandbox
Problem Developers wanted to run AI agents continuously and switch devices without losing state.
Solution They added a Vercel Sandbox plugin that runs agents in the cloud, saves files and conversation history, and reuses the AI gateway key across sessions.
Read on vercel.comDetails -
VercelNewAgent platforms
Box Mount now runs in Vercel Sandbox
Problem Agents needed a way to read and write Box files without exposing credentials.
Solution Created Box Mount that syncs a Box folder to a Vercel Sandbox directory, letting agents use normal file commands while keeping credentials separate.
Read on vercel.comDetails -
GitLabAgent platformsExtend GitLab Duo Agent Platform into your business systems
Problem Engineers had to manually trace errors in business systems, causing delays for users.
Solution Created a flow that lets Salesforce call a GitLab Duo Agent Platform routine via the Flows API, automatically pulling code and context from GitLab to diagnose errors.
Read on about.gitlab.comDetails -
Fly.ioAgent platformsFeng Shui for Robots: The Fly MCP Server
Problem Agents needed a safe sandbox that could easily move to full production infrastructure.
Solution Fly created the MCP server that lets agents run in isolated Sprites, install packages, start services, get HTTPS URLs, checkpoint state, and migrate to production seamlessly.
Read on fly.ioDetails -
Hugging FaceAgent platforms
Turning Database Grading into RL Rewards: Post-Training Agents with ThinkingBox
Problem Agents often said they finished tasks even when the database still needed work.
Solution They created ThinkingBox, a system that runs agents against a real database, checks the final state, and uses that check as a reward signal to fine‑tune the agent.
Success rate rose from 51.70% to 60.78%
Read on huggingface.coDetails -
ShopifyAgent platforms
How Shopify turns open models into task experts
Problem Token‑heavy tasks cost too much and run slowly on the main model.
Solution Created specialized sub‑agents by post‑training open models with a continual learning loop and an LLM judge, then swapped them in for the main model.
7× faster, 90% cheaper
Read on shopify.engineeringDetails -
NVIDIAAgent platforms
How NVIDIA OpenShell Reference Architecture Secures and Governs AI Agent Workloads at Scale
Problem Autonomous agents could perform unsafe actions when working on long, complex tasks.
Solution They created OpenShell, an open‑source runtime that controls what files, services, and devices an agent can access, adds monitoring, and lets operators block or modify actions.
Read on developer.nvidia.comDetails -
RampAgent platformsBuilding Laminar: Ramp's Managed Agents Platform
Problem Existing agent platforms didn’t meet Ramp’s needs, so they needed a custom solution.
Solution Ramp created Laminar, an internal managed‑agents platform that handles tool calling, memory, sandboxing, and observability for their AI agents.
Read on builders.ramp.comDetails -
VercelAgent platforms
Making every harness durable
Problem Developers needed always‑on coding agents without keeping their laptops running.
Solution They moved the coding‑agent harnesses to Vercel’s cloud, creating long‑lived processes that keep session state in memory and can run for minutes or hours.
Read on vercel.comDetails -
monday.comAgent platformsOne prompt, dozens of API calls: rebuilding the monday.com API edge for AI Agents
Problem AI agents were making many API calls, overloading the API edge and causing high latency and resource use.
Solution They rebuilt the API edge to handle rules and routing inside the router, cutting CPU and memory use and reducing added latency to under 100 ms at p99.
92% less CPU, 80% less memory, 11x lower latency
Read on engineering.monday.comDetails -
MetaAgent platforms
The Thin Orchestrator: How We Made Our AI Agents More Reliable by Giving the Router Less to Do
Problem The main AI agent kept hallucinating because it had to handle all domain knowledge and routing.
Solution They created a thin orchestrator that only sequences tasks and checks reliability, while 14 specialized agents handle each domain, cutting context overload.
Token usage cut up to 65‑fold
Read on engineering.fb.comDetails -
Mozilla.aiAgent platforms
Who searches the web for your agent?
Problem Developers cannot see or control which search engine their AI agents use, risking privacy, cost, and flexibility.
Solution Mozilla.ai built Otari, a middleware that intercepts agents' web_search calls, lets you choose or host your own search backend, and keeps the model separate.
Read on blog.mozilla.aiDetails -
NVIDIAAgent platforms
Optimizing Agent Reliability and Tool Calling Efficiency for Chip Design Using NVIDIA NeMo Gym
Problem AI agents often produce plausible but wrong results, risking errors that spread through chip design workflows.
Solution Used NVIDIA NeMo Gym to test and refine the Siemens Fuse EDA agent, adding verifiers and better tool descriptions, which raised benchmark scores to ~95% and cut agent turns by 66%.
Score ~95%, turns reduced 66%
Read on developer.nvidia.comDetails -
Hugging FaceAgent platforms
Run fast, private agents on-device with LFM2.5-24B-A2B
Problem Need fast, private on‑device agents that can reason and use tools without heavy compute.
Solution Created a 24B‑parameter MoE model with only 2B active parameters per token, chain‑of‑thought output, large 131K context, and tool‑calling ability, enabling cheap continuous on‑device inference.
Leads on five tool‑use benchmarks, three‑times next best on AutomationBench
Read on huggingface.coDetails -
Mozilla.aiAgent platforms
Building the Missing Half
Problem Developers were locked into hidden features of proprietary agent APIs, losing control over tools and data.
Solution Created Otari, a gateway that lets agents use open models while giving developers full control over search, code sandboxing, and persistent memory.
Read on blog.mozilla.aiDetails -
MicrosoftAgent platforms
Agent Lightning v1.0: A 3,500-Line Lightweight Agentic RL Framework for Training Agents with Real Harnesses
Problem Training agents with real-world tools was costly and mismatched the deployed version.
Solution They built Agent Lightning v1.0, a lightweight RL framework that inserts an LLM proxy between the agent and the model, letting the same harness used in production train directly.
Read on microsoft.comDetails -
SalesforceAgent platforms
AI Agents Need Memory: Building Persistent Memory for Agentforce
Problem AI agents forgot previous interactions, making complex tasks hard to automate.
Solution Created a persistent memory layer that stores conversation context and task state, letting agents recall prior steps across sessions.
Read on engineering.salesforce.comDetails -
Hugging FaceAgent platforms
AutoSynthData: Generating Training Data for Enterprise Agents
Problem Enterprises lacked enough training data for agents to handle their specific tools, rules, and data.
Solution They created AutoSynthData, a system that turns agent failures and a stronger teacher model's successes into many realistic tasks, validates them, and builds a curriculum that focuses on remaining weak spots.
Read on huggingface.coDetails -
Booking.comAgent platforms
Introducing AgentHub: Booking.com Platform for Modular Travel AI Agents
Problem Travel teams needed a fast way to create AI agents that could handle many booking tasks.
Solution They built AgentHub, a configuration‑driven platform that lets teams assemble reusable LLM agents, connect them to internal APIs, and run multi‑step workflows.
Read on booking.aiDetails -
ShopifyAgent platforms
ShopGym: Realistic, reproducible sandboxes for shopping agents
Problem Developers struggled to train and test shopping agents because real stores constantly change and bots get blocked.
Solution They built ShopGym, a sandbox that copies live stores into resettable simulations and creates realistic shopping tasks, letting agents be evaluated consistently.
Read on shopify.engineeringDetails -
NVIDIAAgent platforms
Tracing Agent Harness Behavior with NVIDIA NeMo Relay
Problem Developers can’t see why an agent takes extra steps or fails, making it hard to improve efficiency.
Solution They created NeMo Relay, which records every model and tool call, timing, and errors, letting developers trace and compare agent runs to find inefficiencies.
Read on developer.nvidia.comDetails -
HoneycombAgent platforms
Introducing AI Ecosystem: Zoom Out to See Your Whole AI Agent Fleet
Problem Teams running many agents found it hard to see what was wrong across the whole fleet.
Solution They built AI Ecosystem, a view that shows all agents, their calls, costs, and failures in one place, letting users drill into any conversation for details.
Read on honeycomb.ioDetails -
HoneycombAgent platforms
Agents Need Context: Introducing Canvas Connectors, Fleet-wide AI Agent Visibility, and More
Problem Teams run many AI agents and need more context to understand production incidents quickly.
Solution Built Canvas, a live workspace that pulls in external data via connectors (GitHub, PagerDuty, Confluence, Jira) and shows evidence like graphs, enabling collaborative investigation of agent‑driven failures.
Read on honeycomb.ioDetails -
ZalandoAgent platforms
Agentic Platform: Open-Sourcing the Agentic Identity Broker
Problem Agents needed to act for users across services without storing users' access tokens.
Solution Created an open‑source identity broker that records delegation chains and securely exchanges tokens so agents can access third‑party systems on behalf of users.
Read on engineering.zalando.comDetails -
WeaviateAgent platforms
Agent Memory with Engram: A Practical Guide
Problem Agents struggled to keep track of past conversations, leading to repetitive or irrelevant responses.
Solution Created Engram, a managed memory service that stores conversation data, extracts facts into topics, and lets agents search past memories before generating replies.
Read on weaviate.ioDetails -
SalesforceAgent platforms
AI Agent Observability: Making Silent Production Failures Visible and Actionable
Problem Silent failures of AI agents in production went unnoticed, hurting reliability.
Solution The team created Agentforce Health Monitoring, which tracks agent uptime, logs metrics, and alerts ops teams to investigate issues quickly.
Read on engineering.salesforce.comDetails -
NVIDIAAgent platforms
TensorRT Edge-LLM Completes the MLPerf Edge Agentic Benchmark 6.4x Faster on Jetson AGX Thor
Problem Edge devices struggled to run multi‑step AI agents quickly within tight power and memory limits.
Solution Created TensorRT Edge‑LLM that runs a 27B model on Jetson AGX Thor using FP4 quantization, tree‑based multi‑token prediction, and KV‑cache reuse, cutting inference time dramatically.
6.4× faster than reference
Read on developer.nvidia.comDetails -
monday.comAgent platforms
Building a Robust Harness for agent in production: Feed agent case-study
Problem AI agents kept giving wrong or costly results, hurting user experience and budget.
Solution They created a custom harness around LangChain that adds data validation, flow boundaries, and feedback loops, ensuring the feed agent returns accurate, actionable items.
Read on engineering.monday.comDetails -
NVIDIAAgent platforms
Building a Memory-Driven Agent with NVIDIA NemoClaw
Problem The AI needed ongoing context about people, projects, and decisions to act like a useful chief of staff.
Solution They built a memory‑driven agent with NVIDIA NemoClaw that stores a structured self‑model in Markdown and a SQLite ledger, retrieves relevant facts, and governs execution.
Read on developer.nvidia.comDetails -
Fly.ioAgent platforms
Your Agent Speaks MCP. Give It a Computer.
Problem Developers needed cheap, safe, on‑demand computers for running AI agents without overhead.
Solution Fly.io created Sprites, disposable cloud VMs with durable storage that agents can call via MCP, providing progressive tool disclosure and isolation.
Read on fly.ioDetails -
Hugging FaceAgent platforms
How Much Memory Does Your Agent Actually Need?
Problem Agents needed to know how much past‑task memory to keep without hurting cost or performance.
Solution They built a system that distills past actions into guidelines, then injects either the full set or a compact, task‑specific subset at inference, tuning the amount per model size.
+16.1pp task completion, +5% token cost
Read on huggingface.coDetails -
ShopifyAgent platforms
Building an agentic harness that outlasts the model
Problem Security teams needed a fast, reliable way to find and fix code vulnerabilities.
Solution They built a multi‑agent harness that scans code, runs tests, and automatically opens pull requests with fixes.
Read on shopify.engineeringDetails -
HoneycombAgent platforms
What Comes After Observability?
Problem Ops teams spent too much time manually investigating latency spikes and updating dashboards.
Solution They built AI agents that automatically run queries, call tools, and edit boards, handling longer tasks and scaling to millions of runs each month.
2 million agent‑initiated queries per month
Read on honeycomb.ioDetails -
RazorpayAgent platforms
Running Hermes at Razorpay: a Network-isolated, Self-improving “Second Brain” for every Employee
Problem Employees needed personal AI assistants that could run continuously without risking security or data leaks.
Solution Created an isolated Kubernetes platform that gives each employee a dedicated Hermes agent, sandboxed storage, and scheduled jobs, letting the agent learn and act autonomously.
15,000 sessions in eight weeks
Read on engineering.razorpay.comDetails -
Booking.comAgent platforms
From Traditional ML to AI Agents: How Booking.com Scales AI Observability With Arize AI
Problem Complex AI workflows were hard to debug because many steps happen behind the scenes.
Solution They integrated Arize AI to collect telemetry and trace every prompt, model call, retrieval and tool use, giving end‑to‑end visibility for all AI systems.
Read on booking.aiDetails -
AmazonAgent platforms
Capturing token IDs during agentic interactions for better reinforcement learning
Problem RL training needed exact token‑level logs of agent actions, but transcripts lost crucial details.
Solution Created Turnstile, a Rust proxy that sits between the agent harness and the model, capturing every token ID as it’s generated and exporting a clean trajectory for RL trainers.
Read on amazon.scienceDetails -
HoneycombAgent platforms
Instrumenting AI Agents for the Agent Timeline: A Practical OpenTelemetry Guide
Problem Debugging multi-step AI agents in production was hard because failures showed only vague model errors.
Solution They added OpenTelemetry spans with a shared conversation ID to every model, tool, and downstream call, letting Honeycomb render a full timeline of the agent’s actions.
Read on honeycomb.ioDetails -
Hugging FaceAgent platforms
Is it agentic enough? Benchmarking open models on your own tooling
Problem Agents waste time and tokens when APIs and docs are unclear or hard to use.
Solution Created a benchmark harness to measure how agents solve tasks with different tools, then redesigned the transformer CLI and examples, letting agents find solutions with far fewer tokens.
Agents used 1.3–1.8× up to 6× fewer tokens
Read on huggingface.coDetails -
Hugging FaceAgent platforms
Agentic Resource Discovery: Let agents search
Problem Agents had to pre‑install every tool, making it hard to scale to thousands of capabilities.
Solution They defined the ARD specification and built a Discover Tool that lets agents search federated registries for tools and skills at runtime using natural‑language queries.
Read on huggingface.coDetails -
Hugging FaceAgent platforms
The Open Source Community is backing OpenEnv for Agentic RL
Problem Developers lacked a common, open tool to connect agents with diverse environments for training.
Solution Created OpenEnv, an open‑source library that standardizes the interface between any agent, environment, and trainer, letting models use any harness.
Read on huggingface.coDetails -
Fly.ioAgent platforms
Building Agents that Don't Break Themselves
Problem Agents often delete or damage files when running commands, causing failures.
Solution Fly created sandboxed Sprite containers and a persistent loop that isolates agent code, letting agents run safely without breaking themselves.
Read on fly.ioDetails -
GoogleAgent platforms
Unlocking dependable responses with Gemini Enterprise Agent Platform’s Agentic RAG
Problem Enterprise queries often need multiple data sources, causing incomplete answers from single-step retrieval.
Solution Built a multi‑agent RAG system that plans, reasons, and iteratively searches across corpora, confirming enough context before answering.
Accuracy improved up to 34%
Read on research.googleDetails -
IntercomAgent platforms
Operator: A look under the hood
Problem Scaling AI support across thousands of workspaces required reliable, precise actions.
Solution Built a production Agent with over 50 specialized tools and 10 skills, plus robust tooling and infrastructure to fetch data, run queries, and orchestrate actions reliably.
Read on intercom.comDetails -
GoogleAgent platforms
ReasoningBank: Enabling agents to learn from experience
Problem Agents kept repeating the same mistakes because they had no way to learn from past failures.
Solution They built ReasoningBank, a memory system that pulls useful patterns from both successful and failed runs, letting agents adapt during use.
Higher success rates and fewer steps on web browsing and coding tasks
Read on research.googleDetails -
Fly.ioAgent platforms
Run Paperclip AI on a Sprite
Problem Developers wanted an easy way to run a multi‑agent AI system on cheap, isolated hardware.
Solution They provide step‑by‑step commands to install Paperclip AI on a Fly.io Sprite, configure networking, authentication, and expose a public URL.
Read on fly.ioDetails -
DeliverooAgent platformsUnlocking AI innovation through automated governance
Problem Teams faced duplicated effort and complex compliance when scaling AI projects.
Solution Deliveroo created a unified AI Hub and Governance-as-Code platform that standardizes tooling, tracing, and evaluation for safe, fast agentic AI deployment.
Read on deliveroo.engineeringDetails -
ShopifyAgent platforms
Building the Universal Commerce Protocol
Problem Merchants have many different payment, discount, and fulfillment rules, making it hard for AI agents to handle purchases.
Solution They created the Universal Commerce Protocol, a layered, extensible API that lets agents discover merchant capabilities, negotiate terms, and complete transactions across any store.
Read on shopify.engineeringDetails
No write-ups match these filters.