The Integrated Package Wins the Procurement Cycle
Microsoft used the Build 2026 keynote and the KPMG announcement seven days later to ship the integrated harness stack — GA eval, GA governance, an open cross-vendor spec, and a flagship Big-4 production deployment — in a single release window. AWS shipped two of the four components first (AgentCore Evaluations GA on March 31, AgentCore Harness preview on April 22), but has not closed the loop with an open spec or a published Big-4 deployment at comparable scale. Google has not headlined any of the four. The 90 days between June 15 and mid-September is the window AWS and Google have to respond before Q3 RFPs for 2027 deployment close. If neither closes the integrated package by then, Microsoft holds the procurement advantage for 12-18 months.
The May 31 Runtime Commoditized post named the buyer-side harness gap: eval, observability, governance, cost control. The runtimes had converged. The harness around them was still the buyer's problem. Two days later, Microsoft answered every layer of that gap in the Build 2026 keynote. Seven days after the keynote, they paired it with KPMG putting Agent 365 governance across 276,000 professionals in 138 countries, including into audit work product, the highest-bar regulated workflow in financial services. The sequence is the story. The runtime layer was the last commodity. The harness around it is where the procurement cycle is being won.
What Microsoft Actually Shipped at Build
Five things landed simultaneously on June 2 and June 3.
Foundry tracing and evaluations: GA. Per the Microsoft Foundry blog, the eval and tracing surface covers "agents built on LangChain, LangGraph, OpenAI SDK, Microsoft Agent Framework, and any custom framework via OpenTelemetry." Framework-agnostic is the claim Foundry's whole eval pitch rides on. The same blog put OpenTelemetry forward as "the common language underneath." Microsoft is positioning Foundry as the OTel-native console for agent observability, the same way Datadog positioned itself for application observability ten years ago.
Open Trust Stack: open-source release. The Open Trust Stack post introduces two cross-vendor primitives. ASSERT, an open-source eval framework targeted explicitly at the 6-13 million GenAI developer population using LangChain, CrewAI, and LiteLLM — not the Foundry installed base. Agent Control Specification (ACS), a portable YAML standard for safety controls at five lifecycle checkpoints: input, LLM, state, tool execution, output. ACS is positioned "alongside Model Context Protocol and Agent2Agent" as a cross-vendor spec, both of which are Anthropic-originated standards Microsoft is now embracing. Partner endorsements at launch: CrewAI, IBM, Arize AI, KPMG, Zscaler. This is the layer competitors have nothing published against.
ROI for agents: private preview. Worth flagging carefully. The Foundry observability blog headlines "from the first inference call to the ROI dashboard" without qualifying GA status. The Open Trust Stack post lists ROI for agents under private preview. The ROI surface is the most aspirational part of the marketing and the least mature part of the product. If you are writing an RFP off this announcement, treat eval and tracing as production-ready and treat ROI as roadmap.
Agent 365: already GA from May 1. Agent 365 went GA on May 1, 2026 at $15/user/month standalone or $99/user/month inside the new Microsoft 365 E7 bundle, per the Microsoft Security blog. It is an identity and governance control plane for agents. Microsoft's docs claim cross-vendor scope covering AWS Bedrock and Google Cloud agents, which has not been independently verified at primary docs depth — treat as vendor-stated until a customer reference confirms.
Foundry Agent Service hosted agents: GA in 30 days. The managed harness path. AWS shipped the same surface in public preview two months earlier (April 22). Microsoft is shipping it GA at the same Build window that takes evaluations GA.
The shape of the bundle is the news, not any single line item.
What KPMG Actually Bought
The June 9 joint release is the procurement signal worth reading carefully.
KPMG is deploying Microsoft 365 Copilot to 276,000+ professionals across 138 countries and rolling Agent 365 governance across the same workforce. The named workflows are audit, tax, and advisory. The hosting substrate is Foundry Agent Service via the KPMG Workbench platform. The bundled annual cost on the E7 SKU lands near $328 million at list pricing per DigitalApplied's third-party analysis; the governance-only line item is closer to $50 million. Neither number is independently audited and Microsoft and KPMG have not confirmed pricing.
Two things make the deployment more material than the headcount suggests.
First, Scott Flynn, Global Head of Audit at KPMG International, framed the rollout as agents that "enhance real-time analysis, earlier risk identification and deliver deeper insights, strengthening audit quality." That is a regulated-output workflow. KPMG is putting agents into the path of audit work product, which means the firm has cleared Foundry plus Agent 365 against the diligence wall the May 13 post named: auditability, permission scoping, data lineage, exception handling, drift monitoring, vendor concentration, reproducibility, bias and fairness, PII handling, liability allocation. Either KPMG built compensating controls externally (possible, given the firm has its own AI governance practice), or Microsoft's stack passes a Big-4 internal audit committee at a level no competing stack currently does. Both readings are interesting. The first means the diligence work is still buyer-side and the runtime choice is somewhat fungible. The second means the integration is real.
Second, KPMG's name appears on the Agent Control Specification partner list in the Open Trust Stack post. KPMG is endorsing the open spec, deploying the governance plane, and integrating the managed runtime — three commitments in one quarter. Lisa Heneghan, Global Chief Digital Officer at KPMG: "This requires strong foundations in governance, visibility and accountability — key to embedding responsible AI." For a Big-4 firm, governance is the work product. The endorsement is the procurement.
I have spent thirteen years inside financial-services governance functions watching deals like this get approved or killed. The thing the press release does not say but the procurement file definitely does is that someone at KPMG already did the model risk validation, the third-party risk assessment, the data-residency walkthrough across all 138 countries, and the SOC 2 mapping. That work takes 9-18 months on a stack with this much novel surface area. KPMG announcing in June 2026 means the diligence started in mid-2025, which means the procurement decision predated the Build keynote by months. The keynote was the marketing window for a deal that was already done. That is also what makes the signal leading. The procurement bet was placed before the integration was announced.
The Comparative Snapshot at Day 0 of Q3 2026
The honest comparison is layer-by-layer, because the runtimes converged in May and the differentiation lives in the surfaces above them.
| Layer | Microsoft (June 2026) | AWS (June 2026) | Google (June 2026) |
|---|---|---|---|
| Runtime | Foundry Agent Service (GA in 30 days) | AgentCore Runtime (GA Oct 2025) | Antigravity 2.0 + Gemini Enterprise Agent Platform |
| Eval / golden-set | Foundry evaluations (GA June 2 2026) | AgentCore Evaluations (GA March 31 2026) | Vertex model evaluation, no agent-eval headline |
| Observability | Foundry tracing (GA June 2026); Azure Monitor + OTLP | AgentCore Observability via CloudWatch (GA) | Gemini Enterprise governance, no observability headline |
| Cross-framework spec | ASSERT + ACS (open-source, Build 2026) | None published | None published |
| Identity / governance plane | Agent 365 (GA May 1 2026) | IAM + AgentCore Identity | Gemini Enterprise governance |
| Cost / ROI dashboard | ROI for agents (private preview) | AgentCore cost controls in observability | Not headlined |
| Managed harness | Foundry Agent Service (GA in 30 days) | AgentCore harness (public preview) | Antigravity Managed Agents API |
| Reference Big-4 deployment | KPMG 276K (June 9 2026) | None published at comparable scale | None published at comparable scale |
AWS got there on individual components first. AgentCore Evaluations GA was March 31, two months before Microsoft. The managed harness public preview was April 22, six weeks before Microsoft. But AWS has not closed the loop with an open cross-vendor spec or a published Big-4 deployment. Google is the laggard. The SiliconAngle Antigravity write-up noted no evaluation or observability specifics beyond CodeMender's code-security scanning at I/O.
Microsoft is the only one of the three that, as of June 15, has simultaneously released GA eval, GA governance, an open cross-vendor spec, and a flagship Big-4 production deployment. The integrated package is what wins the procurement cycle, not the individual GA flags.
The OpenTelemetry Bet
The Foundry blog's "any custom framework via OpenTelemetry" line is the forward bet under the announcement. If the OpenTelemetry GenAI semantic conventions stabilize through Q3 and Q4 2026, Microsoft is positioned as the OTel-native console for agent observability and the broader ecosystem — Arize, Phoenix, Datadog, LangSmith, Braintrust — sits as either complement or competition depending on how the spec consolidates.
The risk is that the conventions stay experimental. As of June 2026 the GenAI conventions are still stabilizing, with six layers of coverage (LLM client spans, agent spans, events, metrics, orchestration, MCP) at varying maturity. AWS supports OTel but does not lead with it in AgentCore branding. Google has no OTel headline. ACS is positioned alongside MCP and A2A, which are themselves Anthropic-originated standards Microsoft is embracing.
The pattern across both standards moves is consistent. Microsoft is betting that open spec plus strong proprietary console wins the architecture debate, and that the proprietary alternatives (AWS internal CloudWatch, Google internal Vertex) end up looking like accidental lock-in rather than deliberate strategy. That is a real bet, not a durable structural advantage — frame it as a positioning thesis with a six-month proof window.
The 90-Day Procurement Window
Enterprise RFPs for 2027 deployments are being written between July and October 2026. Microsoft shipped the integrated harness stack four months before the typical RFP cutoff for Q1 deployment. AWS and Google have parts of the stack in market. Neither has the full package.
Three scenarios for the next 90 days.
AWS responds by Q3 with AgentCore Harness GA plus a published cross-vendor spec. Plausible. AWS shipped Evaluations GA first, the harness is in active preview, and re:Invent in early December is the natural launch window. The cross-vendor spec is the harder ask. AWS historically owns its primitives and has not published an open standard at MCP or ACS's level of cross-vendor positioning. A Big-4 deployment announcement at comparable scale on AWS substrate would also need to land inside this window to neutralize the KPMG marketing.
Google responds at Cloud Next in October with a Gemini Enterprise eval headline and a governance answer. Plausible if Google decides to engage the procurement conversation directly. As of June 15 they have not. The Gemini Enterprise framing has been platform-and-agent-development-focused, not eval-and-governance-focused. The October window is the structural slot.
Neither responds at integration level by Q4. This is the implication of the June timing. The procurement teams that need to write RFPs for January deployment cannot wait for AWS re:Invent in December or Google Cloud Next in October if their RFP cutoff is September. The integrated package is the only option that ships today. Microsoft locks 12-18 months of cross-cloud agent governance procurement advantage.
The KPMG deployment is the leading indicator for which scenario lands. If Deloitte, PwC, or EY announce comparable Big-4 deployments on AWS or Google substrate by October, the integrated-package thesis weakens. It means the diligence work transfers and the runtime choice is more fungible than the procurement signal suggests. If they announce on Microsoft, or stay quiet, the thesis hardens. Big-4 firms watch each other's diligence outcomes; they do not move on procurement together by accident.
What This Means for Buyers
If you are writing an agentic AI RFP for 2027 deployment between July and October 2026, the integration thesis says Microsoft is the only single-vendor option that closes the harness loop today. If you are committed to AWS or Google for non-agent infrastructure, the implied path is multi-cloud agent governance — Agent 365 over AWS Bedrock or Google Cloud — rather than single-vendor consolidation. Microsoft is selling both paths. The KPMG deployment is a cross-cloud reference for the multi-cloud path; it is also a single-vendor reference for organizations that already run on Microsoft.
Three diligence questions to drive into the RFP regardless of which path you pick.
What is the actual GA status of every component you depend on? ROI for agents is private preview. Foundry Agent Service GA timing is "in 30 days" from Build, which means roughly July 2 if the schedule holds. ACS is open-source but the adoption surface is partner-list-deep, not standard-implementation-deep. Do not let the integrated marketing collapse the maturity tiers.
Where does the buyer-side work actually stop? Eval suite definition, golden-set curation, cost-governor thresholds, trace-replay tooling for your specific workflows — these are still buyer-side regardless of which runtime you pick. The integrated package shortens the substrate-build phase. It does not eliminate the harness-build phase for your specific domain.
Who owns the diligence handoff? If you are in a regulated industry, the procurement gate is model risk management, third-party risk, and data-residency review. KPMG cleared all three at full-firm scale in 9-18 months of internal work. If your timeline is "deploy by Q1 2027," start the diligence work now regardless of which substrate you ultimately pick. The runtime decision is downstream of the governance decision.
What This Means for Builders in the Harness Gap
The cross-framework plus OpenTelemetry positioning means Microsoft is trying to consume the substrate that LangSmith, LangFuse, Braintrust, Patronus, and Arize Phoenix sit on. ASSERT specifically targets the 6-13M GenAI developer population that uses LangChain and CrewAI, the same population those tools serve. Arize is on the ACS partner list, which reads as cooperation. CrewAI is on it too. Whether the rest of the observability vendor field reads ASSERT as complement or competitor will determine how the eval-tool consolidation plays out over the next two quarters.
I have been running my own loops on the harness side of this gap for the last six months — small models on local hardware, multi-agent orchestration over Claude Code and Codex CLI, telemetry pipelines through Phoenix and into a SQLite-backed metrics store. The thing the runtime announcements miss is that workload-specific eval is the actual product, not the eval framework. ASSERT lowers the floor for building one. It does not write the eval suite for your specific workflow. The buyer who treats "we adopted Foundry" as the same project as "we have a working harness" is the buyer whose 2027 budget gets burned on uncapped retries by April, exactly the failure mode the pilot-to-production death march named in May.
The procurement cycle is being won at the integration layer. The harness gap is still where the actual engineering happens.