AgentCore Runtime v2 is the next generation of the Amazon Bedrock AgentCore runtime. Sessions restore from a snapshot taken after startup, so they start in seconds and get more compute, but anything your agent creates at startup is shared by every session. Find and move that code before you set the platform version to V2.
What AWS released
The next generation AgentCore runtime was announced by AWS on 12 November 2026 by the general manager of Amazon Bedrock AgentCore. AWS says it is available now in regions across North America, South America, Europe and Asia Pacific. The pricing page and FAQ call it Runtime v2 and keep the earlier runtime as V1.
Access is a setting, not a new service. According to the AgentCore FAQ, you set platformVersion to V2 when you create or update a runtime. An existing agent moves the next time it is updated, unless that update sets the version to V1. You can move back to V1 at any time, and AWS states V1 has no announced deprecation date.
What actually changed
The official pages list these changes:
- Startup: each session restores from a snapshot rather than booting your container. The FAQ gives 1.9 seconds at P75 on a 2 GB image, against 29.7 seconds on the previous version, and the announcement says start time no longer depends on image size or concurrency.
- Bigger sessions: up to 8 vCPUs and 32 GB of memory per session, and container images up to 10 GB.
- Session length: microVM sessions run up to 8 hours without interruption, per the FAQ.
- Isolation: every session still runs in its own hardware-isolated Firecracker microVM, with its own identity, access policy and audit trail.
- Storage: an agent can mount storage such as an Amazon S3 bucket inside the session with FUSE support. The announcement says binding each session to its own storage folder is planned, without a date.
- Billing shape: the pricing page says CPU scales to zero during I/O wait (waiting for the model, a tool or a database), idle memory is reclaimed, and billing is per second. A committed baseline gives a discounted rate for always-on agents.
- Deploys are slower: the FAQ notes deploys and updates now take minutes rather than seconds, because each one builds a new snapshot.
The snapshot catch
This is the part to read before you upgrade. The FAQ says most agents need no changes, but values created at startup, such as random seeds, credentials, timestamps and open connections, are shared by every session restored from the same snapshot.
In practice, look for these patterns in your agent's entry point and module-level code:
- Credentials and tokens fetched at import time. Every session gets the same token, and it may expire mid-way through an 8 hour session. Fetch them per session, or refresh them on use.
- Database, HTTP and model-client connections opened at startup. A restored socket is not a live connection. Open clients lazily inside the session, and make retry logic handle a stale handle.
- Random seeds and IDs. If a session ID, nonce or idempotency key is generated once at startup, sessions will collide. Generate them per request.
- Timestamps. A "started at" value captured at boot will be the snapshot time, which breaks expiry checks and audit records. This matters for the audit trail an agent keeps.
- Per-tenant state. Anything loaded at startup that belongs to one user or tenant now leaks into every session. Re-check this against your multi-tenant data isolation rules.
AWS points to a validator and migration guide in the FAQ. Run them, and also run your own test that starts two sessions and checks they do not share tokens, IDs or connections.
Cost and capacity: what to model
This is CodeDTX's advice, not AWS's. On the pricing page, the V2 consumption rate per vCPU-hour and per GB-hour is higher than V1, while V2 bills less of the session: no CPU while the agent waits on I/O and no memory once it is released. Agents that spend most of their time waiting on model and tool calls are likely to cost less; CPU-heavy agents may cost more. Replay a week of real session traces against both rate cards before you decide, and set budgets as described in controlling AI agent running costs.
The larger session limits are useful for agents that run code, process files or load big local models and tools in the image. Removing warm pools also removes a piece of infrastructure you no longer need to operate.
When not to switch yet
- Agents that redeploy many times a day. Deploys now take minutes, which slows hotfixes and CI pipelines. Plan for that before you move.
- Agents whose startup code you cannot change quickly. If credentials or connections are created at import time across a large codebase, fix that first; the FAQ lets you pin V1 in the meantime.
- Agents that need per-session storage folders. That binding is announced as planned, not available, so keep your current storage scoping until it ships.
Move one low-risk agent first, compare startup, error rates and cost against V1 using your production monitoring, then move the rest.
Frequently asked questions
Do existing AgentCore agents move to Runtime v2 automatically?
Not straight away. According to the AgentCore FAQ, an existing agent moves to V2 the next time it is updated, unless the update sets the platform version to V1. Agents you leave alone stay on V1. That means a routine deploy can switch an agent without anyone deciding to, so pin the version explicitly in your infrastructure code before your next release.
What code breaks under the snapshot model?
Anything created once at startup and expected to be unique or fresh per session. AWS names random seeds, credentials, timestamps and open connections. Common examples are tokens fetched at import time, database pools opened in the module scope, and IDs generated at boot. Move these into the session handler or create them lazily on first use, then test with two sessions running side by side.
Is Runtime v2 cheaper than V1?
It depends on how your agent spends its time. The published V2 rates per vCPU-hour and GB-hour are higher than V1, but V2 does not bill CPU during I/O wait and reclaims memory the agent releases. Agents that mostly wait on model and tool calls should gain, while compute-heavy agents may not. Model it with your own session traces before switching.
Can we move back to V1 if something goes wrong?
Yes. The FAQ says you can set the platform version back to V1 at any time, and V1 has no announced deprecation date. Keep the V1 setting ready in your deployment configuration as a rollback path, and remember that each update builds a new snapshot, so a rollback deploy will also take minutes rather than seconds.



