AgentCore Gateway and MCP 2026-07-28: A UK Upgrade Checklist
MCP 2026-07-28 makes the protocol stateless and HTTP-native. Amazon Bedrock AgentCore Gateway can advertise the new version with one UpdateGateway call — here is the UK enterprise checklist before you flip it.

MCP 2026-07-28 makes the protocol stateless and HTTP-native. Amazon Bedrock AgentCore Gateway can advertise the new version with one UpdateGateway call — here is the UK enterprise checklist before you flip it.
- MCP 2026-07-28 is opt-in and dual-version: AgentCore Gateway adds it via UpdateGateway without recreating targets.
- Sessions and initialize handshakes are gone — tool calls are self-contained; app state must move to explicit IDs.
- Audit clients for logging/setLevel, -32002 error matching, Roots/Sampling/Logging deps, and elicitation support.
- UK rollout: inventory → dual advertise → migrate SDKs → verify OTel/HTTP monitoring → drop old versions.
The biggest MCP revision since launch — without a big-bang cutover
The Model Context Protocol 2026-07-28 specification is the largest change since MCP shipped: the protocol becomes stateless, scales on ordinary HTTP infrastructure, hardens OAuth/OIDC alignment, and adds a governed extensions system so future growth does not require another breaking core bump.
For teams on Amazon Bedrock AgentCore Gateway, AWS's message is deliberately calm: adding 2026-07-28 is a configuration change. You do not recreate the gateway or rewire every Lambda/OpenAPI target. Clients pick a version per request. That is the right shape for UK estates that cannot freeze production agents for a weekend migration.
What actually changed (enterprise view)
| Theme | Before | 2026-07-28 |
|---|---|---|
| Sessions | initialize handshake + Mcp-Session-Id, sticky LB or shared store | Self-contained requests; version/capabilities in _meta |
| Discovery | Tied to session lifecycle | server/discover anytime |
| Intermediaries | Parse JSON-RPC body to route/throttle | Mcp-Method / Mcp-Name headers |
| Caching | Long-lived SSE for catalog invalidation | ttlMs / cacheScope on list/read responses |
| Server→client mid-call | Persistent SSE for elicitation/sampling | Multi Round-Trip Requests with opaque requestState |
| Errors | Often HTTP 200 + in-body error | Real HTTP statuses for transport failures |
| Auth story | Gateway IAM/OAuth unchanged by protocol bump | Still unchanged — upgrade does not rotate credentials |
AgentCore Gateway continues to aggregate Lambda, APIs and MCP server targets behind one MCP endpoint. On the new path, the first tool call is one request, not handshake-then-call.
Why UK teams should care this quarter
- Horizontal scale — Sticky sessions and shared session stores were the tax of early MCP. Stateless MCP fits ALB/API Gateway patterns UK platforms already operate.
- Observability — W3C Trace Context in
_metaplus honest HTTP status codes means CloudWatch/OpenTelemetry dashboards stop guessing. - Regulated change control — Dual-version advertising lets you show CAB a reversible, staged plan instead of a hard cutover.
- Client diversity — Claude Code, Cursor, custom Strands/LangGraph agents and partner tools will upgrade on different clocks. Gateway can speak both until they catch up.
Pre-flight audit (do this before UpdateGateway)
Ask three questions AWS itself recommends, made concrete for UK ops:
- Are clients ready? Confirm MCP client SDKs in each agent framework ship
2026-07-28support. Tier-1 SDKs were expected during the RC window — verify your pinned versions in CI, do not assume. - What breaks if sessions disappear? Search your code and runbooks for: protocol sessions used as app state,
logging/setLevel, literal-32002matching, Roots / Sampling / Logging dependencies, and tools that need elicitation or sampling (clients must advertise matching capabilities or get-32021). - Do you control both sides? You do not need to. Dual-version gateways serve old and new clients simultaneously.
Also note: if a gateway fronts an MCP target that upgrades before clients do, Gateway can translate ordinary tool calls — but elicitation/sampling across that version gap is not supported. Plan those tools carefully.
Safe UK rollout sequence
Stage A — Inventory (3–5 days)
- Catalogue every Gateway ID,
supportedVersions, inbound auth (IAM SigV4 vs OAuth/JWT), and downstream targets. - List agents and SDK versions that call each gateway.
- Flag tools that use elicitation, sampling, progress tokens or custom header-bound fields.
Stage B — Dual advertise (1 change window)
Read current config, then send the complete supportedVersions list (UpdateGateway replaces, it does not append):
aws bedrock-agentcore-control get-gateway \
--gateway-identifier <gateway-id>
aws bedrock-agentcore-control update-gateway \
--name <gateway-name> \
--role-arn <gateway-role-arn> \
--protocol-type MCP \
--authorizer-type <gateway-authorizer-type> \
--gateway-identifier <gateway-id> \
--protocol-configuration '{
"mcp": {
"supportedVersions": ["2025-11-25", "2026-07-28"]
}
}'
Supported versions on Gateway today include 2025-03-26, 2025-06-18, 2025-11-25 and 2026-07-28. Keep whatever you already advertise plus the new one.
Stage C — Prove the new path
Call the gateway with MCP-Protocol-Version: 2026-07-28 and Mcp-Method: tools/list. Success means the version is live. -32022 means it is not in supportedVersions yet (body lists what is).
Migrate pilot agents first (non-prod → one production workflow). Watch HTTP 4xx rates separately from application JSON-RPC errors.
Stage D — Trim only when clean
When inventory shows no clients still requesting the old version, UpdateGateway again without that version. Have a rollback list ready (re-add the old version) before you trim.
Operational gotchas worth a runbook line
- Requests that omit
MCP-Protocol-Versiondefault to2025-03-26. If that default is not insupportedVersions, the call is rejected — so do not drop2025-03-26until every client sends an explicit header. - Header/body contradictions on header-bound fields return
HTTP 400/-32020. - Deprecations (Roots, Sampling, Logging) remain functional for a grace window, but new UK builds should not take fresh dependencies on them — prefer tool parameters, provider APIs and OpenTelemetry.
Bottom line
MCP 2026-07-28 is designed to be the last core break. AgentCore Gateway lets UK estates adopt it the way they should adopt any protocol change: advertise dual, migrate clients, verify telemetry, then retire. The teams that rush to a single-version production gateway will invent an outage. The teams that treat supportedVersions as a controlled rollout list will barely notice the cutover.
Advertise 2026-07-28 alongside your current version first. Dual-version Gateway is the safe UK pattern — migrate clients on your schedule, then trim the old protocol only after inventory says zero sticky-session dependents remain.