AgentCore Runtime V2 GA: A UK SRE Playbook for Elastic Memory and Cold Starts

What shipped on 18 September 2026

AWS announced general availability of the next-generation Amazon Bedrock AgentCore Runtime — the serverless microVM compute layer for production agents. Set platformVersion to V2 when you create or update a runtime.

Two headline fixes for UK SREs who already run agents in production:

1. Elastic memory — sessions start small; memory pages in on demand; unused memory is reclaimed during the session, so you stop paying the peak watermark until session end.
2. Consistent cold starts — the platform snapshots a healthy environment once; new instances restore that snapshot. In AWS testing, P75 cold start stayed ~1.9–2.0 seconds for images from 200 MB to 2 GB, versus roughly 5.4–30 seconds on V1 as image size grew.

Regions at GA include eu-west-1 (Ireland) alongside us-east-1, us-east-2, us-west-2, and ap-northeast-1 — so UK latency-sensitive estates can stay in-region without waiting for a later EU wave.

Why UK production agents felt the V1 pain

Agents moved past chat. Coding agents run for minutes; ambient agents stay event-driven for hours. Under V1:

• Memory tracked the high watermark for the whole session — expensive for bursty UK batch and overnight reconciliation agents.
• Cold starts grew with image size and concurrency — worst under the exact burst traffic that opens Monday morning queues.
• Teams compensated with warm pools and home-grown snapshot tricks — undifferentiated heavy lifting.

V2 targets that gap without abandoning serverless, session isolation, scale-to-zero, or consumption billing.

How V2 works (SRE-relevant)

AWS measured cold starts with an empty echo agent (no model, no tools) across Regions over the public internet — so published P75 numbers include cross-Region RTT. Treat them as platform start path, not end-to-end user latency. In that test the agent body itself was ~34 ms at P75; production wall-clock is still dominated by the agent loop and model calls.

UK production gate before flipping platformVersion: V2

1. Region check — confirm your AgentCore Runtime Region is on the GA list (eu-west-1 for most UK primary estates).
2. Baseline V1 metrics — capture P50/P75/P99 time-to-first-token or time-to-ready, session GB-hours, and error rate for 7 days.
3. Non-prod dual-run — deploy the same agent image on V1 and V2 under a canary alias; compare cold-start and cost, not just happy-path answers.
4. Init hygiene — move heavy one-time loads into startup that completes before healthy; anything lazy that must run every cold start still costs users.
5. Interactive UX tip — open the session when the user enters chat (before they submit), so restore finishes while they type.
6. Observability — keep your existing AgentCore / CloudWatch dashboards; add a panel for cold-start vs in-session latency so regressions are obvious.
7. Rollback — document how to pin platformVersion back to V1 (or previous) if cost or latency regresses for a specific agent class.
8. Security unchanged expectation — hardware-enforced session isolation remains; do not weaken Identity / Policy / Gateway gates because compute got cheaper.

What V2 is not

• Not a replacement for AgentCore Gateway, Identity, or Policy — those are still your tool and auth planes (see prior AIATS Gateway and Policy playbooks).
• Not a guarantee that a 30-second agent loop becomes 2 seconds — only the platform start portion is being flattened.
• Not available in every commercial Region on day one — check the GA Region list before promising UK multi-Region active-active.
• Not a license to ship multi-GB images without review — smaller images still help pull times elsewhere in the pipeline.

Coming soon (plan, do not block on)

AWS called out roadmap items: committed baseline discounts for steady sessions, larger compute/storage, x86 microVMs, suspend/resume with memory snapshotting, and scoped identity for unattended agents. Useful for architecture reviews — not a reason to delay a V2 canary if you already pay V1 peak-memory tax in eu-west-1.

14-day UK cutover sketch

Days 1–3: Inventory runtimes; tag interactive vs long-running; pull cost and latency baselines.
Days 4–7: V2 canary on one interactive agent in eu-west-1; compare P75 ready-time and GB-hours.
Days 8–11: Expand to one long-running / ambient agent; validate reclaim behaviour under idle gaps.
Days 12–14: CAB note with before/after; promote remaining agents by class; keep one V1 control agent for a further week.

Ali's take

V2 is the AgentCore runtime UK SREs asked for after the first production winter: pay for the work, not the watermark, and stop gambling cold starts on image size. Flip the version flag behind a canary, keep Gateway and Policy as hard gates, and measure platform start separately from model time — or you will "prove" V2 failed when the LLM was the slow part all along.