Tuesday's L3+L4 focus surfaces three simultaneous signals from the same April 22 week. Google announced the Gemini Enterprise Agent Platform at Cloud Next 2026 (April 22–24) absorbing Vertex AI (L4+L3); OpenAI launched Workspace Agents on April 22 to ChatGPT Business, Enterprise, Edu, and Teachers as research preview (L4+L3); and AWS shipped Agent Registry preview on April 9 plus AgentCore CLI (14 regions) + Managed Harness + Filesystem Persistence on April 22 (L3+L4). Microsoft Copilot Agent Mode entered Office GA on the same day, completing the four-way structure.
Google Gemini Enterprise GA — Vertex AI Absorbed, A2A v1.0 Default, 200-Model Garden
Google announced the Gemini Enterprise Agent Platform at Cloud Next 2026 (April 22–24, Las Vegas), retiring Vertex AI five years after launch and absorbing it together with Agentspace into a single product. Two components — Gemini Enterprise Agent Platform (the developer surface that succeeds Vertex AI) and Gemini Enterprise App (the team-facing surface for discovering, creating, sharing, and running agents). The platform ships A2A protocol v1.0 as its default interoperability layer and bundles Workspace Studio, a no-code builder, with a 200-model Model Garden that includes Anthropic Claude Opus 4.7.
Pricing moves from per-token to per-agent plus per-seat — Business edition starts at roughly $21 per user per month on annual commit, Standard and Plus tiers run $30–60 per user per month, with 10–20% volume discounts above 500 seats. Samsung SDS announced a strategic partnership with Google Cloud at the same event (Lee Ho-jun, SDS / Ruth Sun, Google Cloud Korea), and LG CNS won the Google Cloud Partner of the Year 2026 award for Korea.
The pricing pivot is the decisive structural change — moving from per-token (which CFOs found unforecastable) to per-seat / per-agent gives procurement teams a forecastable line item that maps to existing software license budgets for the first time.
OpenAI Workspace Agents — Custom GPTs Successor, 60+ Apps, MCP Server Support, Free Through May 6
OpenAI launched Workspace Agents on April 22 as a research preview to ChatGPT Business ($20 per user per month), Enterprise, Edu, and Teachers plans. Six structural changes — (1) Codex-powered cloud agents that run continuously rather than per-chat; (2) RBAC, approval workflows, memory, and analytics built in for organizational control; (3) direct integration with Slack, Salesforce, Gmail, Google Drive, SharePoint, Notion, Atlassian Rovo, and 60+ enterprise apps; (4) skills, files, and custom MCP server support; (5) workspace directory and link-based sharing for team-level agent assets; (6) recurring schedules inside ChatGPT plus direct invocation from connected Slack channels.
Pricing is free through May 6, then credit-based — credit cost per run is undisclosed. Custom GPTs are functionally deprecated. Differentiating levers — leverage of the 700M+ ChatGPT consumer base as the workflow entry point and explicit adoption of Anthropic's MCP standard for custom tool servers. Withholding credit pricing until May 6 hints at internal calibration against ChatGPT Plus ($20) cannibalization risk. Samsung SDS OpenAI Korea reseller (~$370M) channel timing is the key Korean watch.
AWS Agent Registry Preview + AgentCore CLI / Managed Harness — Governance Control Plane Emerges
AWS shipped two waves on top of Bedrock AgentCore. Wave 1 (April 9): Agent Registry preview — three access modes (AgentCore Console UI, AWS CLI/SDK APIs, MCP server interface), IAM plus OAuth (Custom JWT) authentication, semantic and keyword search, approval workflows, CloudTrail audit trails, URL-based auto-discovery that pulls metadata from a live MCP endpoint, in 5 compute regions. Wave 2 (April 22): AgentCore CLI in 14 AWS Regions at no additional charge — deploys agents with infrastructure-as-code (AWS CDK today, Terraform coming); Managed Harness in preview — define an agent by model + system prompt + tools and AWS handles the full reasoning, tool selection, action execution, and response streaming loop; Filesystem Persistence in preview — externalizes local session state so agents can suspend mid-task and resume.
The differentiator is that Bedrock AgentCore is positioned not as an orchestration plane but as a control plane — Anthropic, OpenAI, Cohere, Mistral, and arbitrary frameworks can plug in with their own MCP servers and orchestration logic, while AWS sits in the audit, IAM, and approval seat. The combined message is that "agent sprawl" is now an addressable enterprise problem, immediately aligned with EU AI Act and Korea Framework Act compliance infrastructure.
Lock-in Change — MCP, A2A, AP2 Protocol Stack as Default Assumption
The MCP standard moved one tier closer to inevitability. With Google Gemini Enterprise shipping A2A v1.0 as the default interoperability layer, OpenAI explicitly enabling custom MCP servers in Workspace Agents, and AWS Agent Registry treating MCP servers as first-class catalogable resources alongside agents and tools, the protocol stack — MCP for tool access, A2A for agent-to-agent messaging, AP2 for agent-to-merchant payments — is now the default assumption rather than an Anthropic-specific bet.
Switching costs flow upward from the model layer (where switching is technically easy) to the orchestration plus governance layer (where Registry-style audit trails, approval workflows, and per-seat pricing make migration politically expensive). The single-vendor lock-in frame is being replaced by per-cloud governance lock-in.
Korea Market Implications — SI Dual-Track + Framework Act Alignment
Samsung SDS holds both Google Cloud strategic partnership and OpenAI Korea reseller (~$370M) simultaneously. LG CNS won Google Cloud Partner of the Year 2026 for Korea. SK C&C, NHN Cloud, and Naver Cloud face decision pressure before Q2 quarterly close on which Tier 1 to align with. The Korea Framework Act (effective January 22, 2026) entered its calibration phase in April with a 40-expert public-private task force. AWS Agent Registry's governance primitives may become the model for follow-up MSIT/KISA guidelines.
Naver shut down Clova X on April 9 and pivoted to a Vertical AI strategy with Agent N (search, places, maps, healthcare, finance) launching in 1H 2026. Kakao is refining the Kanana on-device model with full agent integration in KakaoTalk and Kanana Search planned for 1H 2026. Both companies face the same decision variable — whether to ship a domestic gateway before June or prioritize alignment with US hyperscalers.
6-Month Implications: The agent platform layer has split into two tiers. Tier 1 (orchestration and runtime) — Google Gemini Enterprise, OpenAI Workspace Agents, Microsoft Copilot Agent Mode, AWS AgentCore — competes on price model, first-party app integration depth, and bundled frontier model. Tier 2 (governance and audit) — currently led by AWS Agent Registry, with EU AI Act and Korea Framework Act compliance infrastructure as the second purchase every Tier 1 customer must make. Expect the Tier 2 governance category to attract Series A through C capital at accelerating valuations through Q3 2026. [HIGH on the two-tier split, MEDIUM-HIGH on valuation and ASP trajectory specifics]