Next-Gen AgentCore Runtime: A UK SRE Playbook for Scale-to-Zero, Baseline Pricing and S3 Mounts
AWS's next-generation AgentCore runtime brings 8 vCPU/32 GB sessions, sub-2-second starts, scale-to-zero, baseline pricing and S3 mounts. Here is how UK SRE teams should migrate safely.

AWS's next-generation AgentCore runtime brings 8 vCPU/32 GB sessions, sub-2-second starts, scale-to-zero, baseline pricing and S3 mounts. Here is how UK SRE teams should migrate safely.
- AWS announced the next-generation AgentCore runtime on 12 November 2026.
- Sessions now scale to 8 vCPUs and 32 GB, with container images up to 10 GB.
- Sessions start in under 2 seconds, so warm pools can be retired in stages.
- Idle agents scale to zero; committed baseline pricing suits always-on agents with 4x burst.
- Agents can mount S3 via FUSE: govern it as a production data path with KMS and lifecycle rules.
- Per-session storage scoping is promised by re:Invent; until then design for shared agent storage.
- Re-baseline cold start, cost per call and memory before and after migration.
- Confirm eu-west-2 / eu-west-1 availability in the Region table before planning cutover.
What changed
On 12 November 2026 AWS announced the next generation AgentCore runtime, the managed compute layer of Amazon Bedrock AgentCore. The core guarantee is unchanged: every session runs in its own hardware-isolated Firecracker microVM with its own identity, policy and audit trail. What is new:
- Bigger sessions: up to 8 vCPUs and 32 GB per session (4x the previous version) and container images up to 10 GB.
- Faster starts: sessions start in under 2 seconds, independent of image size and concurrency; AWS quotes up to 15x faster starts and says warm pools are no longer needed.
- Pay less for idle: memory is reclaimed within minutes of the agent releasing it and idle agents scale to zero, with up to 68% less memory per session for multi-turn agents.
- Committed baseline pricing for always-on agents: a per-session memory baseline billed as a minimum, with burst up to 4x the baseline, at AWS Lambda MicroVM rates.
- Isolation extends to data: agents can mount and save work to a filesystem such as an S3 bucket via FUSE.
- Coming by re:Invent: per-session scoping that binds each session to its own storage folder through session-labelled credentials.
- Availability is stated as "AWS Regions across North America, South America, Europe and Asia Pacific". Check the AgentCore Region table for eu-west-2 (London) and eu-west-1 (Ireland) before you plan a cutover.
Why UK SRE and platform teams should care
If you followed the warm-pool and cold-start advice of the last year, some of that work is now redundant. If you were holding back coding, browser or local-inference agents because 2 vCPU sessions were too small, that blocker has moved. And the finance conversation changes: idle agents scaling to zero plus a baseline option gives you two distinct cost models to choose from per workload.
The new risk is the S3 mount. A session that can write straight to a bucket is a data path, and for UK GDPR and FCA-regulated estates, data paths need owners, prefixes and retention.
The SRE playbook
1. Re-baseline before you migrate
Run the same workload on the current and next-generation runtime and record p50/p95 cold start, cost per invocation and memory per session. Early customer quotes (35% faster cold starts and 17% lower cost per call in one case; around 60% lower memory footprint in another) are encouraging, but they are not your numbers. Your CAB will want yours.
2. Retire warm pools deliberately
If you built keep-warm schedulers or pre-provisioned capacity, do not delete them on day one. Turn them down in steps behind a flag, watch p95 start latency and error rates, then decommission. Remove the cost line from your FinOps report so savings are visible.
3. Choose a billing model per agent
- Bursty, multi-turn, human-in-the-loop agents: stay on scale-to-zero; they gain most from memory reclamation.
- Always-on agents (queue workers, monitoring agents, long-running orchestrators): model committed baseline pricing and size the baseline at observed p50 memory, relying on the 4x burst for peaks.
Tag every runtime with an owner and cost centre so the choice is reviewable.
4. Size up only where it pays
8 vCPU / 32 GB sessions are useful for codebase compilation, browser automation and local model inference. They are also an easy way to quadruple a bill. Set per-agent maximums in your deployment templates and alert when a session runs at the ceiling for long periods.
5. Govern the S3 mount like production storage
- One bucket (or at minimum one prefix) per agent, never a shared "agent-scratch" bucket.
- SSE-KMS with a customer-managed key, Block Public Access on, bucket policy limited to the runtime execution role.
- Lifecycle rules for working files; agree retention with your DPO if sessions handle personal data.
- Keep the bucket in the same UK or EU Region as the runtime to avoid an accidental cross-border data path in your DPIA.
- Until per-session scoping lands, assume sessions of the same agent can see each other's files, and design accordingly.
6. Shrink the image anyway
10 GB images are now supported and start time is independent of size, but large images still cost you in build time, vulnerability surface and registry storage. Keep model weights and datasets out of the image where you can, and scan images in your pipeline.
7. Update runbooks and alarms
Cold-start alarms tuned for the old runtime will be too loose. Re-set thresholds, add alarms for sessions at max memory and for unexpected S3 write volume, and update incident runbooks to reference the new session limits.
A sensible cutover order
- Non-production research and coding agents (biggest gain from larger sessions).
- Internal multi-turn assistants (biggest gain from scale-to-zero).
- Customer-facing agents, once your own latency and cost numbers are signed off.
- Anything using S3 mounts with personal data, only after the DPIA is updated.
Bottom line
The next-generation AgentCore runtime removes three long-standing excuses: cold starts, small sessions and paying for idle. UK teams should take the savings, but treat the new S3 mount as a production data path from day one, and wait for per-session scoping before putting shared, sensitive workloads behind it.
This release quietly deletes a year of cold-start workarounds. The trap for UK teams is the S3 mount: it is a new data path, and it needs a DPIA line, a KMS key and an owner before it needs a feature flag.


