AWS and Google Cloud spent the past week redrawing the map for enterprise AI agents, and the two moves land almost on top of each other. On August 21, 2026, Amazon Web Services pushed Web Search on Amazon Bedrock AgentCore to general availability, a managed tool that lets AI agents pull live, cited web knowledge without any data leaving a customer’s AWS account. Days earlier, Google Cloud confirmed that its Gemini Enterprise Agent Platform, unveiled at Cloud Next 2026, had folded Vertex AI and Agentspace into a single agent-building stack running on more than 200 models. Together, the launches mark the moment cloud vendors stopped selling AI as a feature bolted onto compute and started selling it as the control plane itself.
For engineering teams, the timing is not a coincidence. Enterprises spent 2025 wiring chatbots to internal data. In 2026 they are wiring autonomous agents to the open web, to each other, and to production systems that used to require a human in the loop. That shift changes procurement conversations, security reviews, and cloud spend forecasts all at once. This piece breaks down what actually shipped, what it costs, who is already using it, and where the two platforms are headed next.
What AWS Shipped: Web Search on Bedrock AgentCore
AWS announced general availability of Web Search on Amazon Bedrock AgentCore as a fully managed, server-side tool. In the announcement, AWS described it as a way “to ground responses in current, cited web knowledge with zero data egress from customer’s secured AWS environment,” a distinction that matters for regulated industries that have avoided sending prompts to third-party search APIs (AWS, official announcement).
The mechanics are straightforward. An agent built on Bedrock AgentCore can now call a web search action alongside its existing tool set, retrieve ranked and cited results, and fold them into its reasoning chain without the developer standing up a separate search integration, API key, or data pipeline. AWS confirmed the feature is “generally available today in the US East (N. Virginia) Region” (AWS, official announcement), with broader regional rollout following the pattern AWS has used for prior Bedrock features: launch in us-east-1, then expand.
Web Search on AgentCore is built to work with OpenAI’s GPT-5.4, GPT-5.5, and GPT-5.6 model family, including the Sol, Terra, and Luna variants that AWS added to Bedrock earlier this year. AWS also rolled out cross-region inference for GPT-5.6 models, extending Sol, Terra, and Luna availability to more than 25 AWS regions, which lets globally distributed teams route agent traffic to the nearest region instead of concentrating every request in Virginia. That matters for latency-sensitive agent workloads where a round trip to us-east-1 from Singapore or Frankfurt adds real delay to a multi-step reasoning chain.
The web search launch did not arrive alone. AWS also added real-time vector search to Amazon DynamoDB, giving developers a native, high-recall vector database with single-digit-millisecond latency instead of forcing them to bolt on a separate vector store like Pinecone or OpenSearch for retrieval-augmented generation. And AWS extended Amazon Quick’s agentic capabilities into AWS GovCloud (US-West), a FedRAMP-authorized region, bringing the “AI teammate” concept to government and defense contractors who cannot use commercial regions. Quick’s agentic AI is now live in eight AWS regions.
What Google Shipped: Gemini Enterprise Agent Platform
Google’s answer came out of Cloud Next 2026, where the company consolidated its scattered AI tooling into one product. Google Cloud framed it plainly: “Today, we’re launching Gemini Enterprise Agent Platform — our new, comprehensive platform to build, scale, govern, and optimize agents” (Google Cloud, official announcement). The platform’s own documentation calls it “our open and comprehensive platform that empowers businesses to rapidly build, scale, govern, and optimize enterprise-grade agents grounded in your enterprise data” (Google Cloud, official documentation).
What that means in practice: Vertex AI and Agentspace, previously separate products with overlapping capabilities, now sit under one roof. The platform ships with a no-code agent builder aimed at Google Workspace users, a developer console exposing more than 200 models (including third-party options like Anthropic’s Claude models, notable for a company usually pushing its own Gemini line first), and a web-browsing agent called Project Mariner for live grounding against the open internet. Google also built in managed MCP (Model Context Protocol) servers and a production version of the Agent2Agent protocol, aimed at letting agents built on different vendors’ stacks talk to each other instead of getting locked into a single-vendor agent silo.
On the model side, Gemini 3.7 Flash reached general availability on August 13, 2026, tuned for low-cost, high-throughput inference inside the agent platform. A week later, on August 21, Google added Grok 4.6 in preview to its Model Garden alongside AlphaFold 3 GA access for approved commercial research users, broadening the platform beyond generic chat-style agents into scientific and specialized reasoning use cases. Google also confirmed Gemini 3.6 Flash, aimed at long-horizon multi-step automation, can cut token consumption by up to 65% on extended agent tasks, and introduced Gemini Spark, a persistent cloud agent that keeps working even after the user’s device goes offline, bundled into a $99.99-per-month AI Ultra subscription tier.
Side-by-Side: Bedrock AgentCore vs. Gemini Enterprise Agent Platform
Both platforms are chasing the same customer: an enterprise that wants to deploy autonomous agents without building its own orchestration, grounding, and governance layer from scratch. But they arrive from different starting points. AWS is extending an existing product, Bedrock AgentCore, with a narrower but immediately useful capability. Google restructured its entire AI portfolio around agents as the primary product, not an add-on.
| Capability | AWS Bedrock AgentCore (Web Search) | Google Gemini Enterprise Agent Platform |
|---|---|---|
| GA status | General availability, August 21, 2026 | Launched at Cloud Next 2026; Gemini 3.7 Flash GA Aug. 13, 2026 |
| Initial region | US East (N. Virginia); cross-region inference for GPT-5.6 spans 25+ regions | Global rollout tied to standard Google Cloud region footprint |
| Supported models | OpenAI GPT-5.4, GPT-5.5, GPT-5.6 (Sol, Terra, Luna variants) | 200+ models including Gemini 3.7 Flash, Grok 4.6 (preview), third-party models like Claude |
| Web grounding | Managed server-side web search tool, zero data egress | Project Mariner browsing agent for live web grounding |
| Agent interoperability | Native to Bedrock AgentCore tool-calling framework | Managed MCP servers plus production Agent2Agent protocol |
| Government/regulated access | Amazon Quick agentic AI live in AWS GovCloud (US-West), FedRAMP-authorized | No GovCloud-equivalent agent platform confirmed as of Aug. 2026 |
| Vector/data layer | DynamoDB native real-time vector search, single-digit ms latency | BigQuery/SAP Business Data Cloud Connect, Cortex Framework v7 GA |
| Notable pricing signal | Usage-based Bedrock pricing (no new agent-specific tier disclosed) | Gemini Spark persistent agent bundled in $99.99/mo AI Ultra plan |
The practical takeaway for architects: AWS is betting that enterprises already standardized on Bedrock will adopt web search as an incremental tool call, not a platform migration. Google is betting that enterprises still deciding on an agent framework will pick the vendor offering the broadest model catalog and the most explicit cross-vendor interoperability story, even if that means adopting a brand-new console.
Azure’s Quieter August Update
Microsoft did not launch a rival agent platform this month, but its August 14, 2026 update touched the infrastructure that agentic workloads depend on. AKS control plane metrics via Prometheus reached general availability, giving Kubernetes operators native observability into cluster health without a third-party exporter. ExpressRoute Resiliency Guard entered preview, aimed at flagging single points of failure in private network connections before they cause an outage. Azure Firewall Premium throughput doubled to 22 Gbps, and Azure Front Door picked up atomic batch rule updates (GA) and mutual TLS support in preview.
None of that is agent-branded, but it is the plumbing agent-heavy workloads need: more network resilience, more firewall throughput, and better cluster visibility as traffic patterns shift from predictable human request volumes to bursty, machine-generated agent traffic. Azure’s approach this cycle looks less like a product launch and more like quietly reinforcing the foundation while AWS and Google fight over the agent layer sitting on top of it.
Enterprise Data Layer: DynamoDB, BigQuery, and the Vector Race
Every agent platform has the same bottleneck: retrieval. An agent that can reason but cannot fetch the right internal document, ticket, or record is not very useful. AWS addressed this by adding real-time vector search directly to DynamoDB, avoiding the need to stand up and sync a separate vector database for retrieval-augmented generation. For teams already running DynamoDB as their primary datastore, that removes a whole category of data pipeline work and a whole category of eventual-consistency bugs between the source of truth and the vector index.
Google took a parallel but distinct path, layering SAP Business Data Cloud Connect for BigQuery and Cortex Framework v7 GA on top of its existing data warehouse, along with Apigee AI Gateway guidance for governance, security, and token-cost control on AI APIs. The instinct is the same across both vendors: stop treating the vector store as a bolt-on and start treating it as a first-class feature of the primary database or warehouse product the customer already pays for. Expect this pattern to keep spreading. Any cloud vendor still selling a standalone vector database as its main retrieval story is going to lose deals to a competitor offering it built into a product the customer already operates.
Who Is Actually Deploying These Agents
Vendor announcements are one thing; production deployments are another. Ryanair confirmed a five-year partnership with Google Cloud to roll out Gemini Enterprise across the airline’s operations, connecting organizational data, automating workflows, and building custom agents. For a company running one of the largest and most cost-sensitive flight networks in Europe, an agent platform is only worth the switching cost if it can touch scheduling, customer service, or crew logistics without breaking compliance requirements airlines operate under. The partnership is one of the more concrete signals that agentic AI on public cloud has moved past pilot programs into multi-year operational commitments.
On the AWS side, the GovCloud extension for Amazon Quick’s agentic capabilities points to a different buyer: federal agencies and defense contractors who need FedRAMP authorization before they can touch a commercial AI product at all. That is a slower-moving but stickier customer base once they commit, because switching clouds after a FedRAMP authorization process is a multi-year undertaking most agencies will not repeat casually.
Historical Context: From Chatbots to Agent Platforms
It is worth remembering how fast this moved. Bedrock launched in 2023 as a managed model-hosting layer, essentially an API marketplace for foundation models. Vertex AI followed a similar path at Google. Through 2024, both platforms were primarily about giving developers a menu of models to call, with retrieval-augmented generation and light agent frameworks bolted on as afterthoughts.
2025 was the transition year. AWS introduced Bedrock AgentCore itself, and Google folded early agent tooling into Agentspace and Vertex AI Agent Builder. Both were still positioned as features within a broader model-hosting product. What changed in the past few months is the reversal of that hierarchy: the agent platform is now the product, and model access is a feature within it. Gemini Enterprise Agent Platform absorbing Vertex AI and Agentspace is the clearest evidence of that reversal on Google’s side. AWS naming its new capability after AgentCore rather than after Bedrock itself signals a similar shift in emphasis, even if the underlying infrastructure has not been renamed yet.
This mirrors a pattern cloud vendors have run before: compute became a commodity, so vendors moved up the stack to managed databases; managed databases became table stakes, so vendors moved up again to serverless and managed Kubernetes. Agent platforms are the next rung, and the vendor that owns the orchestration layer captures the customer relationship even if the underlying model call routes through a competitor’s silicon.
Cost and FinOps Implications
Agent workloads do not bill like traditional compute. A single user request can trigger a chain of model calls, tool invocations, and web searches, each metered separately. That makes cost forecasting harder for FinOps teams used to predictable per-instance or per-request billing. Google’s disclosure that Gemini 3.6 Flash can cut token use by up to 65% on long-horizon tasks is as much a cost message as a performance one: token consumption is the line item finance teams are watching most closely as agent adoption scales, because a single runaway agent loop can burn through a budget that a traditional API endpoint never could.
AWS’s zero-data-egress framing for Bedrock Web Search also has a cost dimension beyond compliance. Data egress fees have historically been one of the more unpredictable line items on an AWS bill, and a web search feature that keeps traffic inside the customer’s AWS environment removes one more variable from the estimate. Whether that translates into meaningfully lower bills depends heavily on how often agents actually invoke the search tool, which is not yet public data.
| Update | Vendor | Date | Status |
|---|---|---|---|
| Web Search on Bedrock AgentCore | AWS | August 21, 2026 | GA (US East N. Virginia) |
| Cross-region inference, GPT-5.6 (Sol/Terra/Luna) | AWS | August 2026 | GA, 25+ regions |
| DynamoDB real-time vector search | AWS | August 2026 | GA |
| Amazon Quick agentic AI in GovCloud (US-West) | AWS | August 2026 | GA, FedRAMP-authorized |
| Gemini Enterprise Agent Platform | Google Cloud | Cloud Next 2026 | Launched |
| Gemini 3.7 Flash | Google Cloud | August 13, 2026 | GA |
| Grok 4.6 in Model Garden | Google Cloud | August 21, 2026 | Preview |
| AlphaFold 3 | Google Cloud | August 2026 | GA (approved commercial research) |
| AKS control plane metrics via Prometheus | Microsoft Azure | August 14, 2026 | GA |
| ExpressRoute Resiliency Guard | Microsoft Azure | August 14, 2026 | Preview |
Security and Governance Considerations
Giving an autonomous agent the ability to browse the live web changes its threat model. A grounding tool that pulls in external content can also pull in prompt injection payloads hidden in a webpage, a risk security teams have flagged since the first retrieval-augmented agents shipped. AWS’s zero-egress design addresses where the data goes, not what comes back in through the search results, so teams adopting Bedrock Web Search still need output filtering and citation verification before agent responses reach production decisions.
Google’s Apigee AI Gateway guidance, aimed at governance, security, and token-cost control for AI APIs, is a direct response to the same concern from a different angle: rate-limiting and monitoring agent-to-API traffic before it causes a cost spike or a data leak. Expect both vendors to keep shipping governance tooling alongside capability launches, because enterprise security teams have made clear that agent adoption stalls without an audit trail for what an agent did, why, and on whose authority.
Market Impact and Competitive Pressure
The near-simultaneous timing of these launches is not a coincidence, it is competitive positioning ahead of enterprise budget cycles that typically lock in Q4. Every enterprise architecture team currently evaluating an agent platform now has a clearer decision matrix: pick AWS if the organization is already deep in the Bedrock ecosystem and wants an incremental capability add, or pick Google if the organization is starting from scratch and wants the broadest model catalog with built-in cross-vendor interoperability through Agent2Agent and MCP.
Azure’s absence from the agent-platform headlines this cycle does not mean Microsoft is out of the race. Copilot Studio and Azure AI Foundry remain Microsoft’s agent-building answer, and the August infrastructure updates (Prometheus metrics, firewall throughput, Front Door mTLS) suggest Microsoft is investing in the layer beneath the agent platform rather than announcing a new one this month. That is consistent with Microsoft’s pattern of iterating quietly on existing Copilot branding rather than launching new product names, but it does mean Azure customers evaluating agent platforms this quarter have less new material to compare against AWS and Google’s headline launches.
Predictions: Where Agent Platforms Go Next
Based on the pace and direction of the past three months of launches, here is where the agent platform race is likely headed:
- Web search expands beyond us-east-1 fast. AWS has followed this exact rollout pattern with prior Bedrock features, launching in Virginia first and expanding within weeks. Expect broader regional coverage for Bedrock Web Search before the end of Q3 2026.
- Vector-native databases become the default, not the exception. DynamoDB’s move mirrors what BigQuery and Cortex Framework are already doing on Google’s side. Standalone vector database vendors will face pressure to differentiate on performance rather than convenience.
- Cross-vendor agent interoperability becomes a selling point, not a technical footnote. Google’s early push on Agent2Agent and managed MCP servers puts pressure on AWS and Microsoft to publish comparable interoperability commitments rather than pushing single-vendor lock-in.
- Token-cost control tooling matures into a standard product category. Both Apigee AI Gateway and the Gemini 3.6 Flash token-reduction messaging point to FinOps-for-agents becoming its own toolchain, not an afterthought bolted onto existing cost dashboards.
- GovCloud and sovereign-region agent access becomes a competitive battleground. AWS’s GovCloud extension for Amazon Quick is a first mover in a category Google and Microsoft will likely need to match for regulated and public-sector customers over the next two to three quarters.
What This Means for Engineering Teams Right Now
Teams currently building on Bedrock AgentCore should test Web Search against existing retrieval pipelines before ripping anything out. A managed tool that removes a maintenance burden is worth adopting even if it does not immediately replace every custom integration. Teams evaluating Google’s Gemini Enterprise Agent Platform for the first time should weigh the migration cost of consolidating Vertex AI and Agentspace workloads against the benefit of a unified console, particularly if the organization already has agents scattered across both products.
For teams on Azure, the August infrastructure updates are worth adopting regardless of which agent platform gets chosen eventually, since better network resilience and cluster observability benefit any workload, agentic or not. And for teams still deciding which cloud to build agent infrastructure on at all, the honest answer is that neither AWS nor Google has a decisive lead yet. AWS wins on zero-egress compliance framing and GovCloud access; Google wins on model catalog breadth and cross-vendor interoperability. The right choice still depends more on where an organization’s data and existing cloud contracts already sit than on which vendor shipped the flashier keynote.
Frequently Asked Questions
What is Web Search on Amazon Bedrock AgentCore?
It is a fully managed, server-side tool that lets AI agents built on Bedrock AgentCore retrieve current, cited web knowledge without any data leaving the customer’s AWS environment. AWS took it to general availability on August 21, 2026, initially in the US East (N. Virginia) region.
What models does Bedrock Web Search support?
OpenAI’s GPT-5.4, GPT-5.5, and GPT-5.6 model family, including the Sol, Terra, and Luna variants that AWS added to Bedrock. Cross-region inference for GPT-5.6 now spans more than 25 AWS regions.
What is the Gemini Enterprise Agent Platform?
Google Cloud’s unified agent-building platform, launched at Cloud Next 2026, which consolidates Vertex AI and Agentspace into one product. It gives developers access to more than 200 models, a no-code agent builder for Google Workspace, and a web-browsing agent called Project Mariner.
Is Gemini 3.7 Flash available now?
Yes. Gemini 3.7 Flash reached general availability on August 13, 2026, and is positioned within the Gemini Enterprise Agent Platform for fast, cost-efficient production inference.
Does AWS or Google offer agent access for government customers?
AWS currently has the clearer offering: Amazon Quick’s agentic AI capabilities are live in AWS GovCloud (US-West), a FedRAMP-authorized region, and available in eight AWS regions overall. Google has not confirmed an equivalent sovereign or GovCloud-specific agent platform as of August 2026.
How does DynamoDB’s new vector search feature work?
Amazon DynamoDB now supports real-time vector search natively, with single-digit-millisecond latency, letting developers run retrieval-augmented generation workloads without standing up a separate vector database and syncing it with their primary datastore.
What did Azure release in August 2026 for cloud infrastructure?
On August 14, 2026, Azure shipped general availability for AKS control plane metrics via Prometheus, doubled Azure Firewall Premium throughput to 22 Gbps, and moved ExpressRoute Resiliency Guard and Azure Front Door mutual TLS into preview.
Which enterprises have publicly committed to these agent platforms?
Ryanair confirmed a five-year partnership with Google Cloud to deploy Gemini Enterprise across its operations, connecting organizational data and automating workflows. That is one of the most concrete public commitments to an agent platform disclosed so far in 2026.
Related Coverage
- AWS Bedrock AgentCore Hits GA, Cuts AI Costs 80% [2026]
- AWS vs GCP vs Azure: GCP Cuts SQL Costs 30% [2026]
- ECS vs EKS: $0 vs $438/Mo Control Plane [2026]
- AWS Interconnect vs ExpressRoute: $0 vs $436/Mo [2026]
- Azure Kubernetes Security GA as Backlog Hits $678B [2026]
- Cloudflare Gives AI Agents a Wallet: 20+ Launches [2026]
For more cloud computing coverage, visit the Cloud Computing section on shattered.io.
