The Agent Enterprise Stack Got Assembled in 48 Hours and Most Teams Will Miss the Architecture

Between June 3 and June 5, 2026, six things happened that, taken individually, looked like vendor product news. Taken together, they describe the assembly of the first complete enterprise agent stack — five layers, owned by different companies, now interoperating in production. Almost nobody is naming the architecture that just became obvious. I am going to name it, explain why each layer matters, and tell you which teams are going to be on the wrong side of it by Q4.

Let me lay out the receipts.

On June 3, Coralogix closed a $200 million Series F at a $1.6 billion valuation to scale its AI-native observability platform for agent workloads — the company's second nine-figure round in eleven months, and the largest dedicated agent-monitoring raise on record. On June 4, Microsoft announced Scout, an autonomous agent for enterprise task execution, and shipped the Foundry Toolkit for VS Code with hosted agents and desktop controls on the same day. Also on June 4, Microsoft released a separate agent governance toolkit covering trust, identity, and compliance for AI agents. Also on June 4, IBM and Google Cloud announced a multi-billion-dollar strategic partnership to scale agentic AI on Gemini Enterprise, with IBM's consulting arm leading deployment. Also on June 4, Cognizant expanded its collaboration with Snowflake to ship Cortex-powered intelligent agents across data engineering and analytics. On June 3, Meta rolled its Business Agent out globally to handle customer interactions for businesses on Facebook, Instagram, and WhatsApp.

Six announcements. Three days. Five layers of one stack. Let me show you what got built and why the architecture is now obvious to anyone willing to look at it directly.

The Five Layers That Just Became Real

The enterprise agent stack has been a theoretical concept for two years. People have talked about "the agentic stack" the same way people talked about "the modern data stack" in 2018 — everybody nods, nobody agrees on the layers, every vendor draws the diagram differently. The June 3-5 announcements are the first time the layers have been occupied by named products from named vendors, deployed at production scale, in the same week. The stack is:

**Layer 1 — Agent Authoring and Runtime.** The surface where developers build, host, and execute agents. Microsoft's Foundry Toolkit for VS Code and the hosted agents it shipped on June 4 are the clearest example: an IDE-integrated surface for designing agents, running them locally or in a managed runtime, and shipping them to production. This is the layer Cursor, Windsurf, and the IDE-sidebar tools have been trying to own. Microsoft just made it a first-class product surface on the most widely deployed developer tool on the planet.

**Layer 2 — Agent Orchestration and Distribution.** The surface where agents get exposed to users, business systems, and other agents. Scout, Microsoft's June 4 announcement, sits here. So does Meta's Business Agent, which Meta rolled out globally on June 3 to let every business "show up for every customer, as if they had an infinite team behind them," per the company's own positioning. These are not chat products. They are agent surfaces — autonomous systems embedded into enterprise workflows, customer service queues, and product surfaces that take actions on behalf of users without per-step confirmation.

**Layer 3 — Agent Identity, Trust, and Governance.** The surface where agents get authenticated, authorized, audited, and constrained. Microsoft shipped a separate agent governance toolkit on June 4 covering trust, identity, and compliance. This is the layer that most enterprise AI deployments have been missing entirely, and it is the layer that is going to determine which agent platforms get to operate inside regulated industries. The Microsoft toolkit covers agent identity issuance, action-level authorization, audit logging, and compliance attestation — the four primitives that turn a model calling tools into a system a CISO can sign off on.

**Layer 4 — Agent Observability and Telemetry.** The surface where agent behavior in production is monitored, traced, alerted, and used to improve the system. Coralogix's $200M Series F, announced June 3, is the largest dedicated bet on this layer to date. The pitch is straightforward: agents fail in production in ways that traditional APM tools cannot see, and the company that builds the observability layer for autonomous systems will own the same position Datadog and New Relic own for traditional services. Coralogix is making the bet that agent observability is a $10B+ market by 2028. At a $1.6B valuation, the venture market is signaling it agrees.

**Layer 5 — Agent Deployment, Consulting, and System Integration.** The surface where agents actually get into enterprise production systems. IBM and Google Cloud's June 4 partnership is the clearest example: IBM Consulting will lead agent deployment on Google Cloud's Gemini Enterprise platform, combining Google's model and infrastructure with IBM's global SI footprint. Cognizant's expanded Snowflake partnership, also June 4, is the second example: deep data-engineering integration as a deployment substrate. These are not product launches. They are channel plays — the way agents actually get into the 80% of enterprises that do not have the in-house capability to deploy agentic AI themselves.

That's the stack. Authoring. Orchestration. Governance. Observability. Deployment. Five layers. All five occupied by named products from named vendors, in production, in the same 72-hour window. The architecture is no longer theoretical.

Why The 48-Hour Timing Matters

I want to be specific about why this matters *now*, not later, and not as a general "agentic AI is taking off" argument. The reason is coordination.

When the layers of a stack are occupied by different vendors in isolation, with no apparent coordination, the stack is fragile. Buyers hesitate because they cannot tell which layer will consolidate, which will commoditize, and which will end up as a feature of a larger platform. Vendor lock-in risk is high. Integration risk is high. The whole category sits in a kind of pre-architectural limbo where nobody is willing to bet big.

What changes when the layers get occupied in the same 48-hour window is the buyer's mental model. Suddenly the stack looks like a stack. The questions change from "should I wait for the dust to settle?" to "which vendor is the right partner for which layer, and how do I assemble a coherent architecture across them?" That mental shift is what unlocks enterprise procurement. It is what unlocks the multi-million-dollar agent deployments that have been stuck in pilot purgatory for two years. It is what unlocks the next round of nine-figure raises, the next wave of strategic partnerships, and the next year of agent platform consolidation.

The 48-hour window is the moment the agent enterprise stack became legible to the people who sign the purchase orders. Everything that happens from here is downstream of that legibility shift.

What Is Actually Being Built In Each Layer

Let me get specific about what each announcement actually contains, because the press coverage has been thin and the marketing language has been thick.

**Microsoft Foundry Toolkit for VS Code** is an IDE-integrated agent-authoring surface that ships with hosted agents — managed execution environments where developers can run agents without standing up their own infrastructure. The "desktop controls" component gives agents access to local OS-level operations, the same kind of surface Apple has been blocking and that Microsoft has now built into Windows as a first-class runtime. The toolkit is the same architectural pattern as GitHub's multi-agent orchestration that went GA at Build, but exposed to enterprise developer teams building custom agents rather than to Copilot subscribers. In practice, this is Microsoft's answer to LangGraph, CrewAI, and the open-source agent frameworks: a managed agent-authoring surface on the most widely deployed developer tool on the planet, with deep OS-level integration that the open-source frameworks cannot match.

**Microsoft Scout** is the autonomous agent Microsoft announced on June 4. The positioning is enterprise task execution across business workflows — the same surface Salesforce Agentforce, ServiceNow's Now Assist, and the Google Gemini Enterprise push are all trying to own. Microsoft's advantage, as I have written before, is the same one it had with Windows: the OS-level integration that makes the agent feel native to the user's environment rather than bolted on. Scout is the product manifestation of the Windows Agent Framework (WAF) Microsoft open-sourced at Build 2026. The WAF layer is the framework. Scout is the flagship agent. The two together describe Microsoft's agent platform strategy more clearly than any single product announcement could.

**Microsoft Agent Governance Toolkit** is the layer that has been most under-discussed and that I think will be most consequential. The toolkit covers agent identity, action-level authorization, audit logging, and compliance attestation. In a regulated industry — financial services, healthcare, government — an agent without an identity primitive is a non-starter. An agent that can call a tool without an authorization primitive is a non-starter. An agent whose actions are not auditable is a non-starter. Microsoft's governance toolkit is the first major vendor to ship all four primitives as a coherent product surface. Expect Google, AWS, and Salesforce to follow with competing toolkits within two quarters. The layer is real. The race to own it has started.

**IBM and Google Cloud Gemini Enterprise Partnership** is a multi-billion-dollar strategic collaboration where IBM Consulting will lead agent deployment on Google Cloud's Gemini Enterprise platform. The deal structure combines Google's model and infrastructure with IBM's global system integrator footprint — the same playbook IBM used to dominate enterprise services for mainframes and ERP systems. The strategic significance is that Google just bought IBM's enterprise distribution for Gemini Enterprise. That is a meaningful vote of confidence in Google's agent platform, and a meaningful shift in IBM's strategy away from Watson and toward being a deployment partner for the leading frontier labs. Watson, as a strategic identity, is over. IBM is now the SI partner for Microsoft, Google, and AWS. That is a defensible position. It is not the position IBM spent a decade building Watson to occupy.

**Cognizant-Snowflake Intelligent Agents** is the second SI partnership of the day, narrower in scope, focused on data engineering and analytics. The Cortex-powered agents on Snowflake's data cloud let enterprises build agents that operate directly on the governed data layer. This is a real architecture: agents on the data substrate, with Snowflake's role-based access control as the authorization primitive, Cortex as the model routing layer, and Cognizant as the integration partner. If your enterprise is Snowflake-heavy, this is the path of least resistance for getting agents into production analytics workflows. It is also a shot across the bow of Databricks and Palantir, who are trying to own the same layer from different angles.

**Meta Business Agent Global Rollout** is the only one of the six that targets consumers and small businesses rather than enterprise. The pitch is "every business can show up for every customer, as if they had an infinite team behind them." In practice, this means a small business on Instagram can deploy a Business Agent that handles DMs, comments, and basic customer service queries without human intervention. The strategic significance is not the product — every major lab has shipped a customer-service agent — it is the distribution. Meta controls three of the largest messaging surfaces on the planet (Facebook, Instagram, WhatsApp). When Meta ships an agent product across those surfaces, it bypasses every enterprise procurement bottleneck. Small businesses do not need a CIO, a CISO, or a procurement process. They need a Meta Business account, which most of them already have. The agent stack's distribution layer just got a billion-user surface, overnight.

**Coralogix Series F at $1.6B Valuation** is the largest dedicated bet on the agent observability layer to date. The company raised $115M Series E eleven months ago. The Series F at $1.6B is the second nine-figure round in less than a year, which means the company is either growing faster than its cash burn or the venture market is willing to fund a multi-year cash burn to win the agent observability category. My read: both. Agent observability is a real category with a real technical problem (autonomous systems fail in ways traditional APM cannot trace), a real buyer (every enterprise deploying agents in production), and a real moat (the more agent traffic a vendor sees, the better its anomaly detection gets — a data network effect that compounds over time). Coralogix is the early leader in this category. Expect Datadog, New Relic, Splunk, and Honeycomb to ship agent-specific observability features within two quarters. The layer is real. The race is on.

The Architecture That Just Became Obvious

Put the six announcements together and the enterprise agent stack looks like this:

• **Authoring**: Microsoft Foundry Toolkit, LangChain, CrewAI, AutoGen, Google's Agent Development Kit
• **Orchestration**: Microsoft Scout, Salesforce Agentforce, ServiceNow Now Assist, Google Gemini Enterprise, Meta Business Agent
• **Governance**: Microsoft Agent Governance Toolkit, Auth0 for Agents (Okta), and a wave of governance startups that will get acquired or crushed by Q4
• **Observability**: Coralogix, Datadog (to be announced), Honeycomb, Splunk, Arize (for agent evals)
• **Deployment**: IBM Consulting + Google Cloud, Cognizant + Snowflake, Accenture + Microsoft, Deloitte + AWS, the SI partner ecosystem

Five layers. Each layer has at least one named product, from a named vendor, in production. The architecture is no longer theoretical. It is the structure enterprises will use to plan their agent strategy for the next three years.

Who Is On The Wrong Side Of This

Here is the part of the post where I tell you who is going to lose, because every architectural shift produces losers and pretending otherwise would be dishonest.

**First, the indie agent framework startups that thought they were platform plays.** LangChain, CrewAI, AutoGen, the open-source agent frameworks — they are now authoring-layer tools in a stack where Microsoft has shipped a managed authoring surface on the most widely deployed developer tool on the planet, and where the deployment layer is consolidating around IBM-Google, Cognizant-Snowflake, and the major SIs. The frameworks will survive as developer tools. They will not survive as platforms. The venture-backed ones need to either raise on the strength of authoring-layer metrics and adjust their pitch, or get acquired before the next round. The market is not waiting for them to figure out which layer they own.

**Second, the customer-service-agent startups that thought they were the distribution layer.** Sierra, Decagon, the customer-service-agent cohort — they are now competing with Meta Business Agent, which ships to a billion-user surface for free, and with Microsoft Scout, which ships into every Microsoft 365 enterprise customer with enterprise procurement ready to go. The customer-service-agent category is not dead. The addressable market for "better than Meta's free tier" is shrinking fast. The startups that survive will be the ones that pick a vertical, a regulatory regime, or a data moat that Meta's general-purpose agent cannot replicate. The general-purpose horizontal plays are in trouble.

**Third, the in-house agent governance tools at Fortune 500 companies.** Every large enterprise I have talked to in 2026 has some custom-built agent governance solution — usually a wrapper around Okta, an in-house policy engine, and a lot of bespoke audit logging. Microsoft's governance toolkit, Okta's Auth0 for Agents, and the major cloud providers' competing offerings are about to make those in-house solutions look like rebuilding Windows. The right move for most enterprises is to retire the in-house governance layer, adopt the vendor primitive, and focus the internal engineering team on the domain-specific policy and compliance requirements that the vendor primitives do not cover.

**Fourth, the agent observability roll-your-own solutions.** Teams that have built custom tracing, custom eval, custom alert systems for their agent deployments. The Coralogix Series F is the clearest signal that the agent observability layer is a product category, not a build-it-yourself category. The roll-your-own solutions will look increasingly expensive to maintain and increasingly insufficient in capability. The right move is to evaluate the vendor solutions, pick the one that fits the deployment stack, and stop maintaining the custom one.

**Fifth, the teams that do not have an agent strategy at all.** If your 2026 H2 plan does not include a concrete answer to "which layer of the agent enterprise stack does our team own, and which layers do we buy," you are about to be out-executed by the teams that do. The five layers are real. The vendors occupying them are real. The procurement budgets are real. The only question is whether your team has a position in the architecture or is consuming the architecture from the wrong side.

What I Am Telling You

The agent enterprise stack just got assembled in 48 hours, and the architecture is now obvious to anyone willing to look at it directly. Microsoft owns the authoring and governance layers. Google, Microsoft, Salesforce, and Meta own the orchestration layer. Coralogix, Datadog, and the traditional APM vendors are competing for the observability layer. IBM, Cognizant, and the major SIs own the deployment layer. The framework and protocol layer is still open territory — Anthropic's MCP, Google's A2A, and the open-source A2A implementations are still in flux.

If you are building anything in the agent space in 2026, you need a clear answer to: which layer does this team own, and which layers do we integrate with. The "we are building the whole stack" answer is a death sentence. The "we are building on the stack" answer is the only one that survives the next eighteen months.

The 48-hour window is the moment the architecture became legible. The teams that internalize it now will be the teams defining the next eighteen months of enterprise AI. The teams that miss it will be writing post-mortems about why their carefully built agent platform got out-executed by Microsoft, Google, IBM, and Coralogix — and they will not have a good answer for how they missed all five layers getting occupied in the same week.

That is not a comfortable thing to publish on a Friday. It is also not something you can afford to ignore.

*Sources: Microsoft announcements on Scout, Foundry Toolkit for VS Code, and Agent Governance Toolkit, June 4, 2026; IBM-Google Cloud strategic partnership on Gemini Enterprise, June 4, 2026; Cognizant-Snowflake expansion on Cortex-powered intelligent agents, June 4, 2026; Meta Business Agent global rollout, June 3, 2026 (about.fb.com/news/2026/06/meta-business-agent); Coralogix $200M Series F at $1.6B valuation, announced ~June 3, 2026 (SecurityWeek, VentureBurn, AI Weekly). Agent insider-threat research covered separately.*