Amazon Adds Filesystem Access to S3 After 20 Years, Targeting AI Agent Workloads

Amazon Web Services on Monday launched S3 Files, a feature that turns any S3 bucket into a mountable NFS filesystem. EC2 instances, Lambda functions, containers. All of them can now access S3 data through standard file operations. The service, built on Amazon's Elastic File System, delivers roughly 1-millisecond latencies for active data and supports up to 25,000 simultaneous compute connections per bucket, according to the company's announcement. It shipped in all commercial AWS regions after nine months of customer testing, GeekWire reported.

The feature addresses one of cloud computing's oldest infrastructure frustrations. For nearly two decades, developers who stored data in S3 had to copy it into a separate file system before any application, training pipeline, or AI agent could work with it using standard file operations. That duplication cost money and created synchronization headaches across entire organizations. No more, if AWS's pitch holds.

How the engineering team arrived at its design on the second try

Andy Warfield, the VP and distinguished engineer who leads S3 engineering at AWS, published a candid essay describing how the first internal design attempt failed. The core tension, as Warfield framed it, is structural. Files can be edited in place and shared across applications in real time. Objects in S3 cannot. They are stored and retrieved as complete units, and millions of applications depend on that guarantee.

"We locked a bunch of our most senior engineers in a room and not let them out till they had a plan that they all liked," Warfield wrote. "Passionate and contentious discussions ensued. And then finally we gave up."

The breakthrough, as Warfield described it, came when the team stopped trying to force files and objects to behave identically. S3 Files treats file access as a presentation layer instead. When you mount a bucket, EFS imports metadata from S3 and creates a synchronized view. Files under 128 KB get their data pulled onto high-performance storage automatically. Larger files stay in S3 until you actually read them. Warfield calls it "lazy hydration," and it means you can mount a bucket with millions of objects and start working immediately.

Changes flow back to S3 roughly every 60 seconds as single PUT operations, using what AWS describes as a "stage and commit" model borrowed from Git. If a conflict occurs between filesystem and API writes, S3 wins. The filesystem version gets moved to a lost+found directory with a CloudWatch metric flagging the event.

The pricing math

AWS says S3 Files can deliver over 4 terabytes per second of aggregate read throughput and more than 10 million IOPS per bucket. A feature called "read bypass" routes sequential reads directly to S3 through parallel GET requests, hitting 3 gigabytes per second per client, according to Warfield's essay.

Pricing tells the story of what AWS actually built here. The rates are EFS Performance-optimized, lifted wholesale: $0.30 per GB-month for high-performance storage. Reads cost $0.03 per GB. Writes cost $0.06. That is not a coincidence. S3 Files runs on EFS hardware.

The catch lives in the metering. Every data access operation carries a 32 KB minimum. Read a 1-byte config file? You just paid for 32 KB of throughput. Chief Cloud Economist Corey Quinn warned teams to model their I/O patterns before committing.

AWS claims the feature can cut costs by up to 90% compared to cycling data between S3 and separate file systems. The savings come from the tiering model. Infrequently accessed data drops to S3 Intelligent-Tiering's lower tiers, starting at roughly $0.0125 per GB-month for data untouched for 30 days. S3 Files only charges its surcharge on the fraction you actively touch, and S3 Intelligent-Tiering moves data between tiers for free, unlike EFS, which charges $0.01 to $0.03 per GB for every tier transition.

Why AI agents are the real pitch

AWS's marketing around S3 Files leans heavily on agentic AI. The company positions the feature as eliminating "session state loss and shared-state failures that have blocked multi-agent pipelines in production," VentureBeat reported.

And the pitch tracks with how many agent frameworks work in practice. Tools like LangChain and AutoGen typically read and write local files. They expect filesystem semantics. Asking them to learn S3's object API was a barrier few teams bothered to solve. Now an AI agent can mount a bucket with a single command and work with standard Python libraries and shell scripts. No specialized storage code.

AWS has been layering capabilities onto S3 for months. S3 Vectors hit general availability at re:Invent in December 2025 after a July preview, adding similarity search and RAG against the same bucket data. Stack the two features and S3 starts to look less like a storage service and more like the operating layer for agent infrastructure. That is clearly the bet AWS is making.

Where competitors stand

Google Cloud and Oracle Cloud have nothing comparable, according to storage publication Blocks and Files. Azure Data Lake Storage supports NFS v3.0 and SMB but cannot natively merge file and object access on identical data.

Third-party alternatives exist. Qumulo's Cloud Native platform delivers over 1 TB/s throughput and exceeds 1 million IOPS across all three major clouds. Amazon FSx for NetApp ONTAP runs natively on AWS with S3 access through S3 Access Points.

But S3 Files carries one advantage none of them match, AWS contends. It works directly against existing S3 buckets. No migration. No data copying. No third-party licensing.

AWS says more than 500 trillion objects now sit in S3 worldwide. For those customers, the company describes S3 Files as a filesystem toggle switch where there used to be a multi-week engineering project.

Zuckerberg's Bots Run Meta. Cursor's Bot Ran From Beijing.

San Francisco | Monday, March 23, 2026 Mark Zuckerberg is building an AI agent to help manage Meta. His employees already built their own, and the bots now talk to each other autonomously. One trigge

The Implicator

OpenAI Launches Frontier to Manage AI Agents From Rival Vendors in One System

OpenAI on Thursday released Frontier, a platform that lets companies build, deploy, and manage AI agents from multiple vendors inside a single system. The product works with agents built by OpenAI, by

The Implicator

Moltbook Left Every AI Agent's API Keys in an Open Database, Security Researcher Finds

Anyone could have posted as Andrej Karpathy's AI agent this week. Or as any of the 32,000-plus bots registered on Moltbook, the viral social network for autonomous AI agents. A misconfigured Supabase

The Implicator

Harkaram Grewal

New Delhi

Freelance correspondent reporting on the India-U.S.-Europe AI corridor and how AI models, capital, and policy decisions move across borders. Covers enterprise adoption, supply chains, and AI infrastructure deployment. Based in New Delhi.