# AWS Adds OpenFold3 to SageMaker and Speeds AgentCore V2 for Agent Workloads

> New managed offerings pair open-source biomolecular structure prediction with snapshot-based agent session restarts, expanding AWS's reach into scientific and enterprise AI workloads.

*Published 2026-09-27 · Updated 2026-11-14 · By The Intel Desk*

Amazon Web Services (AWS) is a cloud computing provider that sells on-demand infrastructure and managed services, including machine learning platforms and agent runtime tools.

The two releases target distinct workloads but share a common strategy: Amazon Web Services is pushing managed infrastructure deeper into specialized compute patterns, from scientific inference to agent session orchestration. The first update brings OpenFold3-preview, an open-source biomolecular structure prediction model, to SageMaker JumpStart with asynchronous inference support. The second makes the next-generation AgentCore Runtime the default for Bedrock agents, using snapshot-based session restores to cut cold starts to under two seconds. Both announcements position AWS as an infrastructure layer for workloads that are too specialized for general-purpose compute and too complex for teams to operate themselves.

The OpenFold3 release is part of a broader pattern in which AWS has courted scientific computing with managed services for genomics, drug discovery, and now structure prediction. SageMaker JumpStart, the model catalog inside Amazon SageMaker, has become the entry point for that strategy. The catalog lets researchers deploy a model with a few clicks, then scale it with the same tooling that enterprise teams use for production machine learning. For OpenFold3, AWS chose an asynchronous inference pattern rather than a real-time endpoint, a signal that the expected workload is batch-oriented research rather than interactive use.

OpenFold3-preview was developed by the OpenFold Consortium and the AlQuraishi Lab at Columbia University and is distributed under the Apache 2.0 license, according to an AWS Machine Learning Blog post. The consortium is a collaboration among academic and industry groups working on open-source protein structure prediction, and the AlQuraishi Lab is a Columbia University research group led by computational biologist Mohammed AlQuraishi. The Apache 2.0 license means the model weights and code can be used, modified, and redistributed with few restrictions, which is an unusual level of openness for a field that has been shaped by proprietary systems.

OpenFold3 predicts joint three-dimensional structures for proteins, DNA, RNA, small-molecule ligands specified as SMILES strings, and multi-chain complexes. That all-atom scope is the model's defining feature. Earlier structure prediction tools focused on proteins; OpenFold3 extends the same approach to nucleic acids and ligands in a single prediction run. For drug discovery teams, the capability matters because most therapeutic questions involve how a small molecule binds a protein target, and the interaction cannot be understood by modeling the protein in isolation.

The model writes output in mmCIF format and reports pLDDT, PAE, and PDE confidence metrics, according to AWS. mmCIF is the standard crystallographic file format used by structural biologists, so predictions can be loaded directly into familiar tools. pLDDT is a per-residue confidence score, PAE estimates the positional error between residue pairs, and PDE provides a per-atom error metric for all-atom predictions. Researchers can use those metrics to filter low-confidence regions before investing in follow-up experiments.

## How does the OpenFold3 JumpStart deployment work?

JumpStart is SageMaker's catalog of pretrained models, and OpenFold3-preview appears in it as a deployable model card. Users select an instance type, create an endpoint, and submit structure inputs for prediction. Because inference is asynchronous, jobs are queued rather than executed inline, which lets teams run large batches without holding open a connection. AWS says the deployment relies on SageMaker asynchronous inference endpoints, which queue requests, warm GPUs, and support optional autoscaling and scale-to-zero.

- Asynchronous inference endpoints decouple request submission from result retrieval.
- GPU warmup reduces latency for the first prediction after idle periods.
- Optional autoscaling adjusts endpoint capacity with request volume.
- Scale-to-zero support shuts down idle endpoints to control cost.
- Request queuing lets bursty research workloads submit many jobs without throttling.

The asynchronous pattern fits the economics of GPU inference. Structure prediction is compute-intensive, and a single batch of predictions can occupy a GPU for hours. A synchronous endpoint would force researchers to pay for idle capacity between batches. With request queuing and scale-to-zero, a lab can submit a batch, let the endpoint spin up, run the predictions, and then shut down automatically. The trade-off is latency: results arrive when the queue drains, not in real time. That is acceptable for most research pipelines, which are built around batch processing.

For teams that want to integrate OpenFold3 into an automated pipeline, the output format and confidence metrics are the hooks. A pipeline can parse mmCIF files, filter residues by pLDDT, and rank candidate ligands by predicted binding geometry. The same infrastructure can be shared across a research group, with each project submitting jobs to a common endpoint. AWS's choice to host the model in JumpStart, rather than as a separate service, means it sits alongside other pretrained models in the same console and uses the same IAM, CloudWatch, and billing controls.

The target users are structural biologists, computational chemists, and drug discovery groups that already work in Python and run batch jobs. Many of these teams have access to AWS through academic credits or institutional accounts, but they may not have the DevOps skills to operate a GPU cluster. JumpStart removes that barrier. The model card abstracts away the container image, the inference server, and the autoscaling policy. The researcher supplies inputs and reads outputs, and the platform handles the rest.

## What changed in AgentCore Runtime V2?

Amazon Bedrock AgentCore Runtime V2 is now the default runtime for agent sessions, according to an AWS What's New announcement. The update changes how sessions start. Instead of re-initializing an agent environment from scratch on every session, the runtime restores each session from a snapshot of the initialized agent. That shift cuts startup overhead, makes sessions more ephemeral, and lets teams scale idle agents to zero without paying for warm containers.

AWS says the next-generation runtime starts sessions up to 15 times faster, scales to four times more compute, and runs up to 68 percent more efficiently than the prior version, according to an AWS Machine Learning Blog post. Sessions can use up to 8 vCPUs, 32 GB of memory, and container images up to 10 GB. Each session runs in a hardware-isolated Firecracker microVM. Firecracker is the same virtualization technology AWS built for Lambda, and it is designed to deliver the isolation of a virtual machine with the density of a container.

*AgentCore Runtime V2 specifications per AWS*

| Attribute | Specification | Source |
| --- | --- | --- |
| Session start time | Under 2 seconds P75 cold start on a 2 GB image | AWS What's New |
| Performance gain | Up to 15x faster session starts vs. prior runtime | AWS Machine Learning Blog |
| Compute scale | Up to 8 vCPUs per session | AWS Machine Learning Blog |
| Memory | Up to 32 GB per session | AWS Machine Learning Blog |
| Container image size | Up to 10 GB | AWS Machine Learning Blog |
| Isolation | Firecracker microVM per session | AWS Machine Learning Blog |
| Idle behavior | Scales to zero | AWS What's New |

The snapshot-based approach is the architectural core of the update. Agent sessions often need a warm environment with dependencies installed, credentials configured, and tools loaded. Rebuilding that environment on each interaction is slow and expensive. By restoring from a snapshot, the runtime preserves the initialized state and applies it to a new microVM, which is why cold starts drop below two seconds. The snapshot also enables a clean lifecycle: a session can be created, used, and discarded without leaving state behind.

## What does the AgentCore V2 update mean for enterprise teams?

Coinbase is an early user. In an AWS blog post, Siwei Shen, vice president of engineering at Coinbase, said the runtime lets the company run AI-driven development at scale. Each coding agent session runs in its own hardware-isolated environment, which Shen said the company needs when agents execute code. The runtime scales elastically with demand and starts fast, so engineers move quickly without managing infrastructure.

> AgentCore Runtime lets us run AI-driven development at scale. Each coding agent session runs in its own hardware-isolated environment, which we need when agents execute code, and the runtime scales elastically with demand and starts fast every time, so our engineers move quickly without managing any infrastructure.Siwei Shen, VP of Engineering, Coinbase

The hardware isolation point is worth emphasizing. When an agent executes code, it is effectively running an untrusted program. If the runtime shared environments across sessions, a single bad agent could corrupt another tenant's work or leak data. Firecracker microVMs give each session a separate kernel and device interface, so a failure in one session cannot cross into another. That isolation is the difference between a demo agent and an agent that can be trusted with production credentials.

Scale-to-zero is the cost lever. Agent workloads are often bursty: a team triggers a batch of coding tasks, the agents run for minutes, and then demand stops. Keeping a warm pool of agent environments between batches wastes compute. With scale-to-zero, the runtime releases resources when sessions end and rebuilds them from snapshots when new work arrives. The sub-2-second cold start makes that trade-off painless, because the time cost of starting a session is small enough to ignore.

The snapshot model also changes how developers think about agent state. Instead of assuming an agent persists in memory, teams can design agents that save important state to durable storage and treat the runtime environment as disposable. That pattern is common in serverless computing, and AgentCore V2 extends it to agent sessions. The result is a runtime where the marginal cost of an idle session approaches zero and the marginal cost of a new session is measured in seconds.

## What are the limits teams should consider?

AgentCore V2 sessions cap at 8 vCPUs, 32 GB of memory, and 10 GB container images. Workloads that exceed those limits need different orchestration. The snapshot pattern also assumes the agent environment can be serialized and restored; state that depends on external resources, such as live network connections or local GPU state, may not carry cleanly across a snapshot restore. Teams running agents that hold long-lived connections should test whether the snapshot model fits their workflow.

For OpenFold3, the preview label signals that the model is still under active development. Teams planning production pipelines should validate outputs against their own benchmarks and track upstream releases. The asynchronous endpoint pattern, while well suited to batch jobs, requires a polling or callback mechanism for result delivery, which adds integration work. Research teams without that infrastructure may need to build a small orchestration layer on top of the endpoint.

## What is behind AWS's push into specialized AI workloads?

The two launches are not isolated features; they are part of the same strategic bet. AWS is treating managed AI infrastructure as a horizontal layer with vertical applications. OpenFold3 gives biology teams a path from model card to production inference without GPU cluster management. AgentCore V2 gives software teams a runtime designed for the session lifecycle of agentic workloads. Both are built on the same SageMaker and Bedrock substrate, which may let AWS capture demand from both scientific and enterprise AI buyers.

The common thread is operational automation. AWS is not selling a better model in either case; the model in the first release is external and open source, and the runtime in the second is an infrastructure improvement. The value proposition is that teams can run these workloads with less overhead, which is the same argument AWS has made for managed services since the launch of SageMaker. By extending that argument to biomolecular prediction and agent session orchestration, AWS is trying to make itself the default place to run the next generation of AI workloads.

## What comes next?

The OpenFold3 preview will likely evolve as the consortium releases new versions, and AWS's JumpStart catalog will carry those updates if the model proves popular. AgentCore V2, now the default, will absorb more of the agent workloads running on Bedrock. The more interesting question is whether the two trajectories converge. An agent that can call a biomolecular prediction endpoint is a plausible product: a research assistant that designs experiments, submits structure predictions, and analyzes results. AWS's infrastructure now supports both halves of that workflow.

For buyers, the practical takeaway is that AWS is lowering the cost of entry for two hard categories. Biology teams no longer need a dedicated GPU team to run structure prediction. Software teams no longer need to build custom session orchestration for agents. The caveat is that both offerings are young: OpenFold3 is a preview, and AgentCore V2's snapshot model has yet to prove itself at the largest scale. Teams that pilot now will learn the limits early, but they will also be first to benefit from the cost and latency improvements.

## Sources

1. [OpenFold3-preview can be deployed from Amazon SageMaker JumpStart; the OpenFold Consortium and the AlQuraishi Lab at Columbia University developed OpenFold3 as an open source biomolecular structure prediction model.](https://aws.amazon.com/blogs/machine-learning/openfold3-on-amazon-sagemaker-jumpstart-predict-biomolecular-structures/)
2. [Sessions now start up to 15x faster, scale to 4x more compute, and run up to 68% more efficiently; the new runtime restores each session from a snapshot of your initialized agent.](https://aws.amazon.com/blogs/machine-learning/run-agents-securely-and-efficiently-with-the-next-generation-agentcore-runtime/)
3. [The new runtime restores each session from a snapshot of your initialized agent, scales compute and memory with the work, scales idle agents to zero, and achieves a P75 cold start under 2 seconds on a 2 GB container image.](https://aws.amazon.com/about-aws/whats-new/2026/11/new-agentcore-runtime-default/)
4. [Amazon Web Services launched OpenFold3 for biomolecular structure prediction in SageMaker JumpStart and detailed snapshot-based patterns for AgentCore V2 to accelerate AI agent session startups.](https://tevvio.com/ai-news)

---
Source: https://aiintelreport.com/ai-agents/ai-agents
Index: https://aiintelreport.com/llms.txt · Full text: https://aiintelreport.com/llms-full.txt
