Microsoft Ignite 2025: The seven things that actually matter

Four days in San Francisco. Twenty-plus sessions. Three hands-on labs. Thousands of conference attendees bathed in the blue glow of Microsoft's AI ambitions. Microsoft Ignite 2025 was the most strategically coherent conference Microsoft has delivered in a decade — and that coherence is precisely what should make enterprise engineering leaders pay close attention. Not because the technology is not impressive. It is. But because coherent strategies from dominant vendors create coherent lock-in, and the line between "integrated platform" and "inescapable dependency" has never been thinner.

This is not a session-by-session summary. I have written those individually. This is the synthesis — the big-picture assessment of what Microsoft is actually building, what it means for organisations evaluating these technologies for real deployment, and what was conspicuously absent from four days of announcements.

The seven themes that define Ignite 2025

Across every keynote, breakout, theatre session, and hands-on lab, seven themes emerged repeatedly. Some were stated explicitly. Others required reading between the lines. Together, they paint a picture of Microsoft's strategy for the next three to five years — a strategy that is as much about competitive positioning as it is about technical capability.

1. The agent platform play: Owning the orchestration layer

The thesis Microsoft is betting the company on

If there was a single message from Ignite 2025, it was this: Microsoft intends to own the agent orchestration layer for the enterprise. Not just the models. Not just the infrastructure. The entire stack from agent creation to deployment, governance, and inter-agent communication.

The evidence is overwhelming. Azure AI Foundry provides the model hosting and agent development environment. The unified Microsoft Agent Framework (the Semantic Kernel and AutoGen merger) provides the programming model. Agent 365 provides the governance and registry layer. Work IQ, Fabric IQ, and Foundry IQ provide the business semantics, data preparation, and context engineering layers. The A2A protocol provides agent-to-agent communication. MCP (Model Context Protocol) provides agent-to-tool communication. Agent Factory bundles the lot under "one meter" — unified billing.

This is not a collection of announcements. It is a vertically integrated platform play. Microsoft is building the AWS of AI agents — a platform so comprehensive that the switching costs become prohibitive once you commit.

What the hands-on labs revealed: The unified agent-framework lab was the most telling. Participants built multi-agent systems with A2A communication, MCP tool integration, and Magentic-One orchestration in 75 minutes. The framework genuinely reduces cognitive overhead compared to bolting Semantic Kernel and AutoGen together. The multi-agent Foundry session demonstrated shared state management, human-in-the-loop patterns, and OpenTelemetry observability baked into the runtime. The code works. The abstractions are well-designed. The developer experience is materially better than the two-framework world that preceded it.

The competitive context: AWS Bedrock provides model access and basic orchestration. Google Vertex AI offers models and ML operations. Neither claims to provide the complete business semantics layer that Fabric IQ promises, nor the governance control plane that Agent 365 represents. Microsoft is betting that the platform — not the models — is where enterprise value accrues. Given that models are commoditising rapidly (DeepSeek demonstrated that economics can be disrupted overnight), this is probably the right strategic bet.

The IQ family flywheel: The three IQ components create a self-reinforcing cycle that deserves specific attention. Fabric IQ learns your organisation's "Language of Business" by analysing your Microsoft 365 data. Foundry IQ engineers that understanding into versioned, tested context. Work IQ uses that context to power agents that understand your specific business. The more your organisation uses Microsoft 365, the more data Fabric IQ learns from, the better the context becomes, the more useful the agents are. This flywheel is deliberately sticky — your organisation's semantic map is not portable.

The honest assessment: The platform is impressive and genuinely useful. It is also a flywheel designed to make Microsoft increasingly indispensable. Every agent registered in Agent 365, every semantic map built in Fabric IQ, every context engineering pipeline in Foundry IQ deepens the dependency. That is not a criticism of the technology. It is a description of the business model. Enterprises need to enter this relationship with open eyes about the long-term commitment they are making.

2. Governance as the differentiator: The enterprise unlock

Microsoft's most underrated strategic move

Whilst the keynote focused on capability — what agents can do — the most strategically important sessions were about control — what agents are allowed to do, and how you ensure they do only that.

The AI agent governance session presented a maturity model that should alarm most enterprises: Level 0 (ungoverned) is where the majority sit today. Agents are being built and deployed across business units with no central visibility, no registry, no safety evaluation, and no runtime enforcement. The shadow AI problem is worse than shadow IT ever was, because shadow agents have API access to production data and systems.

The Foundry Control Plane session went further. Sarah Bird opened with KPMG research showing that public trust in AI is declining across every measured dimension — people are more worried, perceive AI as less trustworthy, and are less willing to rely on it. Meanwhile, 81% of business leaders plan to integrate AI agents within 18 months. The gap between leadership enthusiasm and public scepticism is a governance problem, not a technology problem.

The AI red teaming lab made this visceral. Attendees attacked their own deployed models and watched them fail. Automated adversarial testing generated thousands of attack variations — jailbreaks, indirect prompt injection, information disclosure — that human testers would never systematically cover. The gap between "we tested it manually" and "we tested it with automated red teaming" is orders of magnitude.

Why governance is the differentiator: Every cloud provider can host models. Every framework can build agents. But Microsoft is betting that enterprises will choose the platform that makes AI auditable, compliant, and governable. This is the classic Microsoft enterprise play — the same strategy that made Active Directory the backbone of corporate IT for two decades. Own the identity and governance layer, and the compute follows.

Agent 365 is the clearest expression of this strategy. It extends Entra ID to treat agents as first-class identity objects alongside users and devices. Agent registry, access control, behaviour visualisation, security integration with Defender and Purview — this is identity and access management applied to autonomous systems. If your organisation already manages identities through Entra, managing agent identities through Agent 365 is a natural extension. Which is exactly the point.

What was missing from the governance story: Discovery of agents that were built before the governance framework existed. Cross-platform governance for agents not built on Azure. How to handle policy conflicts between business units (compliance says log everything; privacy says some things must not be logged). The cost and friction that governance controls add to development velocity. These are not trivial gaps — they are the gaps where governance programmes actually fail in practice.

3. The identity foundation: Most organisations are not ready

The prerequisite nobody wants to talk about

Here is the uncomfortable truth that Ignite 2025 danced around without confronting directly: most enterprises cannot deploy AI agents at scale because their identity infrastructure is not fit for purpose.

The identity modernisation session made the case that in the AI era, identities are not just about authentication. Systems like Microsoft 365 Copilot depend on identities to understand organisational structure, map how individuals connect to each other, enrich AI responses with context, and enable secure agent interactions. Without accurate, complete, and well-governed identity data, agents cannot determine who should see what, who reports to whom, or which organisational context applies to a given request.

The practical implications are significant:

If your Entra ID has stale group memberships, orphaned accounts, and inaccurate reporting lines — and most enterprise directories do — then AI agents will inherit those inaccuracies. A Copilot agent that uses the organisation chart to route approvals will route to the wrong people. An agent that uses group membership to determine data access will expose data to people who should not see it. Fabric IQ's "Language of Your Business" relies on accurate organisational data to map subject matter experts and decision-making structures. Feed it stale data, and it builds a stale model.

This is not a new problem. Active Directory hygiene has been a standing audit finding in most enterprises for twenty years. But AI amplifies the consequences. A human user with incorrect group membership might stumble into a SharePoint site they should not access. An agent with incorrect permissions will systematically access data it should not, at machine speed, across the entire organisation.

The readiness gap: Microsoft's vision assumes clean identity data, well-defined organisational relationships, and properly configured Conditional Access policies. The reality in most enterprises is a directory that has accumulated two decades of technical debt, where group membership is a best-effort approximation of actual access needs, and where nobody has audited service accounts since the last compliance scare.

Before you deploy AI agents, fix your identity infrastructure. This is unsexy advice, and no vendor will lead with it, but it is the prerequisite for everything else Microsoft announced at Ignite 2025.

4. Operations automation: The tier that gets AI-native first

Where the ROI is calculable and the path is clearest

If there was one area where Ignite 2025 moved beyond aspiration to actionable technology, it was operations automation. The Azure SRE Agent, AKS AI Ops, and the operations-focused labs demonstrated AI capabilities with calculable ROI and clear adoption paths.

Azure SRE Agent is the most pragmatic AI agent announcement from the entire conference. Its pricing model is transparent and usage-based: a baseline of approximately GBP 218 per month for always-on monitoring, plus GBP 68.40 per hour of active incident response. You can calculate whether it saves money before committing. That alone sets it apart from every other agent announcement at Ignite, where pricing ranged from "TBD" to "contact sales."

The SRE Agent detects incidents by correlating telemetry across Azure Monitor, Application Insights, and Log Analytics. It diagnoses root causes by generating KQL queries, cross-referencing deployment history via GitHub integration, and building evidence chains — the same investigative workflow an experienced engineer follows, compressed from 30 minutes to 30 seconds. It remediates across a spectrum from inform-only to autonomous, with guardrails including blast radius limits, rollback capability, and approval gates. Microsoft claims 20,000 engineering hours saved across their own deployments.

AKS AI Ops took this further into Kubernetes territory. The hands-on lab had participants simulate production failures — traffic spikes, memory pressure, cascading pod evictions — and deploy self-healing agents to handle them autonomously. The aks-mcp server, an open-source MCP server for natural-language Kubernetes interaction, was the standout component. Instead of writing complex kubectl pipelines and correlating output mentally during a 2am incident, you ask "what changed in the last 30 minutes that could explain this failure pattern?" and get a synthesised answer with evidence. For experienced operators, this reduces cognitive load during incidents. For less experienced team members, it materially deepens on-call rotation capability.

Why operations is where AI agents land first: The pattern recognition is well-defined. The remediation actions are bounded. The ROI is measurable. The trust can be built incrementally — start in observe-only mode, validate recommendations, enable low-risk automation, and gradually extend scope. This is the same process you would use to onboard a junior engineer to the on-call rota. The mental model is right, and that matters more than any technical specification.

The multi-agent compliance session with SymphonyAI extended this pattern beyond infrastructure into business operations. Their "Always On Compliance" platform uses specialised agents for transaction monitoring, sanctions screening, and fraud detection — replacing the periodic review cycle with continuous, real-time compliance checking. The architecture pattern is the same as SRE Agent: AI handles pattern-based detection at machine speed, humans handle judgment calls and novel situations.

The broader implication: Operations automation is the Trojan horse for the agent platform. Once an engineering team deploys the SRE Agent and sees it handle 70% of routine incidents, the conversation about deploying agents in other business functions becomes much easier. Microsoft knows this. The SRE Agent is not just a product — it is a platform adoption vector.

5. The edge and Physical AI: Extending beyond the cloud

The most ambitious and least production-ready theme

Microsoft's partnership with NVIDIA on Physical AI and the Foundry Local announcements represent the conference's most forward-looking theme — and the one with the widest gap between vision and production reality.

Physical AI — AI models that understand physical properties like geometry, materials, and dynamics — was demonstrated through NVIDIA Omniverse integration with Azure. The thesis: robotics development shifts from programming explicit movements to training AI models in simulation, validating with synthetic data, and deploying through closed-loop digital twin integration. Wandelbots showed commercial products built on this stack, reducing robot programming time from days to hours with hardware-agnostic deployment across ABB, KUKA, Fanuc, and Universal Robots platforms.

The architecture is genuinely impressive. Simulation runs at thousands of times real-time speed on NVIDIA GPUs. Synthetic data generation produces labelled training data in hours that would take months to collect physically. The closed-loop pattern — where real-world performance feeds back into simulation for continuous improvement — means the sim-to-real gap narrows with each deployment cycle rather than remaining a fixed limitation.

But the practical constraints are significant. The stack requires Omniverse licenses, Azure GPU compute, OpenUSD pipeline expertise, simulation engineers, and ML specialists with reinforcement learning experience. The talent required is scarce and expensive. The economics only work for organisations deploying hundreds of robots across complex, variable tasks. And simulation fidelity degrades significantly in unstructured environments — the dusty, wet, thermally extreme conditions found in mining, agriculture, and construction, where I spend most of my professional life.

Foundry Local addresses a different edge problem: running AI inference on personal devices and edge infrastructure. The four use cases — privacy, latency, cost, and offline scenarios — are well-defined and genuine. The PhonePe case study, with on-device financial insights for 600 million users under Indian financial regulations, demonstrates real production value at extraordinary scale. The expansion to Android and Kubernetes edge deployments broadens the addressable market beyond Windows.

The hybrid cloud-edge pattern that emerged across sessions is architecturally significant: route simple queries to local models for instant response, route complex reasoning to cloud models for accuracy, degrade gracefully offline. This requires consistent APIs across cloud and edge (achieved through OpenAI-compatible formats) and intelligent routing logic. It changes how architects think about AI infrastructure — from "which cloud endpoint do I call?" to "where should inference happen for this specific query?"

NPU acceleration on Copilot+ PCs provides the hardware foundation for edge AI. The silicon is shipping. The software stack is catching up. Within two refresh cycles, most enterprise laptops will have dedicated AI compute. The organisations that build edge-aware AI architectures now will be better positioned to exploit that hardware when it is ubiquitous.

6. The competitive landscape: How Microsoft is positioning

Reading between the lines of what was and was not said

Microsoft positioned itself in isolation at Ignite, which is standard conference practice but not how procurement decisions work. Reading between the lines reveals how they are positioning against specific competitors.

Against AWS: Microsoft's integrated platform story — Agent Factory as one meter, one governance framework, one development experience — is a direct counter to AWS's best-of-breed approach with Bedrock. Microsoft is arguing that integration is worth more than flexibility. AWS would counter that their approach avoids lock-in and allows mixing providers. Both arguments have merit, and the right answer depends on your organisation's risk tolerance and multi-cloud reality.

Against Google: Google's strength is in model research (Gemini, DeepMind) and data infrastructure (BigQuery). Microsoft counters with enterprise integration — identity, governance, compliance, and the Microsoft 365 ecosystem. The subtext: Google builds for developers and researchers; Microsoft builds for enterprises and CIOs. This is a simplification, but it drives procurement conversations in boardrooms.

Against Databricks: The Azure Databricks Agent Bricks session revealed the partnership tension that enterprise customers should watch carefully. Databricks is simultaneously a Microsoft partner (running on Azure) and a competitor (offering its own agent development capabilities through Agent Bricks, Lakebase, and serverless workspaces). Microsoft's response is Fabric IQ — a direct play to own the data semantics layer that Databricks considers its territory. The "Language of Your Business" mapping in Fabric IQ competes directly with Unity Catalog's ambitions to be the enterprise data governance standard. If you are using both Azure and Databricks, pay attention to how this competition evolves.

Against the open-source ecosystem: Microsoft's embrace of MCP (originally from Anthropic) and A2A as open protocols is a strategic bet. By adopting open standards, Microsoft reduces the "walled garden" criticism. But the implementations demonstrated — Agent 365 governance, Foundry Control Plane enforcement, aks-mcp server — are Azure-specific. The protocol is open; the production-grade implementation is proprietary. This is the same playbook Microsoft ran with Linux containers on Azure: embrace the standard, differentiate on the implementation, capture the enterprise market.

The model router as geopolitical positioning: Microsoft's announcement of support for Anthropic Claude alongside OpenAI, Meta, Mistral, and a "curated version" of DeepSeek is not just about customer choice. It is Western AI alliance formation. DeepSeek is not offered raw — it is filtered and compliance-wrapped. The message: "Model diversity without geopolitical risk." Whether "curated" provides genuine safeguards or plausible deniability is a question enterprises should ask directly. The economics matter too — if Chinese labs produce competitive models at a fraction of the cost, the pricing pressure on the entire model market intensifies. Microsoft's model-agnostic platform strategy hedges against this by ensuring they capture value at the orchestration layer regardless of which model wins.

7. What was missing: The questions Ignite did not answer

The gaps matter as much as the announcements

Every technology conference is defined as much by what is absent as by what is announced. Ignite 2025 had significant gaps that enterprise decision-makers need to acknowledge.

Pricing transparency across the agent stack

The SRE Agent's transparent, usage-based pricing was the exception that proved the rule. For the rest of the agent platform — Agent Factory, Work IQ, Fabric IQ, Foundry IQ, Agent 365 — pricing was conspicuously absent. "One meter" sounds appealing until you realise you have no idea what that meter will cost. Enterprises cannot build business cases on vision decks. We need numbers. And the absence of numbers this late in the product cycle suggests that Microsoft has not yet resolved the commercial model — which means early adopters are buying risk alongside capability.

Honest lock-in discussion

Not once across four days did any Microsoft presenter acknowledge the vendor lock-in implications of the agent platform. Once your organisational semantics are mapped in Fabric IQ, your agents are registered in Agent 365, your context engineering pipelines are in Foundry IQ, and your governance policies are in the Foundry Control Plane — what is the exit strategy? Can you export your business language maps? Can you migrate agent configurations to a non-Microsoft platform? Can you take your safety evaluations and red-teaming results elsewhere?

These questions were not asked because the conference environment does not encourage them. But they must be asked — and answered — in procurement conversations.

Multi-cloud reality

Microsoft presented its AI stack as if Azure were the only cloud. In reality, most large enterprises operate across Azure, AWS, and GCP. The governance frameworks, agent registries, and observability tools announced at Ignite are Azure-native. Can Agent 365 govern agents deployed on AWS Lambda? Can the Foundry Control Plane enforce guardrails on agents running in Google Cloud Run? As far as Ignite 2025 was concerned, the question does not exist. As far as enterprise architecture is concerned, it is one of the most important questions to answer.

The small and mid-sized enterprise path

Every demo featured large enterprises — BMW achieving 12x faster test data analysis, Epic Systems tracking patient compliance, PhonePe serving 600 million users. Every architecture assumed dedicated platform engineering teams, ML specialists, and governance boards. What is the path for a 500-person company that wants to deploy three agents? The talent, infrastructure, and cost assumptions underlying Microsoft's vision are realistic for Fortune 500 companies. They are aspirational for everyone else. Microsoft 365 Copilot Business was mentioned in the Book of News, but the agent platform story is decidedly enterprise-scale.

Failure modes and limitations

The fictional company "Zavva" used throughout Ignite demos was perfectly compliant, had clean data, and experienced no regulatory friction. It is the AI transformation equivalent of cooking demonstrations with pre-chopped ingredients. Real enterprises have decades of technical debt, contradictory data definitions across departments, regulatory constraints that vary by jurisdiction, and organisational politics that resist algorithmic decision-making.

What happens when agents make mistakes at scale? Who is liable when algorithmic recommendations cause business harm? What are the detailed incident response procedures for agent failures? What can the platform genuinely not do? Microsoft showcased capabilities and avoided discussing boundaries. Every technology has limits; enterprises need to know where they are before they deploy, not after.

The workforce conversation

"Your workforce won't be replaced; it'll be augmented" was the subtext of every human-in-the-loop demo. Whether this framing survives contact with cost-cutting executives seeking headcount reduction remains to be seen. Microsoft has a commercial incentive to position AI as augmentation rather than replacement — replacement reduces the number of Microsoft 365 seats. But the Epic healthcare demo, with its patient compliance scoring, and the SRE Agent's claim of 20,000 engineering hours saved, tell a more complex story about where human labour fits in the AI-native enterprise.

What I would tell my CTO

If I were briefing a CTO after attending Ignite 2025, here is what I would say. Practical recommendations, free of conference excitement and vendor enthusiasm. These are ordered by urgency, not by excitement level — which is precisely why "fix your directory" comes first and "deploy AI agents" comes fourth.

1. Fix your identity infrastructure before deploying agents

This is the unglamorous prerequisite. Audit your Entra ID. Clean up stale group memberships. Verify reporting relationships. Implement proper Conditional Access policies. Audit service accounts. Every AI agent will inherit your directory's inaccuracies, and at machine speed. If your organisation chart in Entra does not reflect reality, neither will your agents.

Timeline: Start immediately. Allow six months for meaningful remediation.

2. Pilot the SRE Agent on a non-critical workload

The Azure SRE Agent has the clearest ROI story of anything announced at Ignite. The pricing is transparent. The capability is narrow and well-defined. The adoption path is incremental — start in observe-only mode, validate recommendations, extend autonomy gradually.

Calculate your specific ROI using actual incident data. If your engineering team spends more than 20 hours per week on operational incidents and your infrastructure runs on Azure with mature observability, the maths likely works. If not, wait for broader availability and GA pricing.

Timeline: Pilot within three months. Evaluate after 90 days of data.

3. Establish an agent governance framework now, not later

Do not wait for the shadow agent crisis. Establish an agent registry. Define registration requirements. Create a governance board with representatives from platform engineering, security, and AI/ML teams. Define safety evaluation criteria and pre-deployment gates.

You do not need Agent 365 to start this. A spreadsheet and a process are better than nothing. But if you are already in the Microsoft ecosystem, evaluate Agent 365 when it reaches general availability.

Timeline: Framework defined within two months. Enforcement within six months.

4. Evaluate the agent-framework for your next agent project

If you are building on Microsoft's stack, the unified agent-framework is the right foundation. It genuinely reduces complexity compared to using Semantic Kernel and AutoGen separately. MCP tool integration is practical. Magentic-One orchestration handles real-world task ambiguity better than static routing.

But do not migrate existing production agents immediately. The framework merged recently. Wait for two to three stable releases before migrating production workloads. New projects should start on agent-framework; existing projects should migrate when the framework demonstrates stability over time.

Timeline: New projects immediately. Migration evaluation in six months.

5. Demand pricing transparency before committing to the agent platform

Do not build business cases on "one meter" without understanding the meter. Before committing to Agent Factory, Work IQ, Fabric IQ, or Foundry IQ, require detailed pricing models from your Microsoft account team. Per-agent pricing, consumption-based pricing, enterprise licensing — the model determines whether the economics work for your organisation's scale.

If Microsoft cannot provide pricing clarity, that is useful information. It means the commercial model is not mature, which means you are buying risk alongside capability.

Timeline: Pricing conversation before any strategic commitment.

6. Build a multi-cloud governance strategy regardless

Even if you are Azure-first, assume you will need to govern agents across multiple platforms eventually. Define governance requirements in vendor-neutral terms. Evaluate whether Microsoft's governance tools can extend to non-Azure deployments. If they cannot, build governance abstraction layers that work across providers.

The worst outcome is discovering your governance framework is platform-locked when a business requirement demands agent deployment on AWS or GCP.

Timeline: Strategy defined within six months.

7. Invest in context engineering as a discipline

Foundry IQ's "Context Engineering" concept — treating the information AI receives as engineered infrastructure rather than ad-hoc prompts — is the most underappreciated announcement from Ignite. Most AI failures are context failures, not model failures. Missing information, irrelevant data, outdated facts. The model is only as good as the context it receives.

Invest in people who understand both business domains and AI capabilities. These "context engineers" will determine whether your agents produce useful results or hallucinated nonsense. Start versioning your prompts, testing context quality, and measuring the impact of context changes on AI performance. This is valuable regardless of whether you adopt Microsoft's specific tooling.

Timeline: Identify and develop internal candidates within six months.

The verdict: Impressive engineering, incomplete answers

Microsoft Ignite 2025 was a technically impressive conference. The engineering is real. The agent-framework works. The governance tools address genuine enterprise needs. The SRE Agent has calculable ROI. The Physical AI vision is architecturally sound. The unified platform story is coherent in a way that competitive offerings are not.

But coherence is not completeness.

The pricing story is largely absent. The lock-in implications are unaddressed. The multi-cloud reality is ignored. The path for organisations without Fortune 500 resources is unclear. The workforce implications are euphemised. The failure modes are undiscussed.

Microsoft is building the operating system for enterprise AI agents. That is not hyperbole — it is an accurate description of the integrated platform spanning identity, governance, orchestration, data semantics, model operations, and deployment infrastructure. If this platform delivers on its promises, it will be as foundational to the next decade of enterprise computing as Windows Server and Active Directory were to the last two.

The risk is the same risk that Active Directory created: indispensable infrastructure that becomes impossible to replace. The organisations that adopted Active Directory early gained genuine competitive advantages in IT operations. They also spent the next twenty years unable to leave Microsoft's ecosystem. Some would argue they never left.

The AI agent platform play is the same pattern at higher stakes. The capabilities are more powerful. The lock-in is more comprehensive. The switching costs will be more punitive. The business value, if the technology delivers, will be more significant.

My advice: adopt deliberately. Deploy the components with clear ROI — the SRE Agent, the agent-framework for new projects, the governance tools. Resist the temptation to commit to the entire platform before pricing, portability, and failure modes are understood. Fix your identity infrastructure first, because everything Microsoft announced at Ignite 2025 depends on it working properly. And above all, build governance before capability, because the most dangerous outcome is deploying powerful AI agents into an organisation that is not ready to control them.

The conference promised to put intelligence back into AI. The platform architecture delivers on that promise. Whether wisdom came along for the ride is a question each organisation will have to answer for itself.

What to watch in 2026

MCP ecosystem adoption. If MCP becomes the universal standard for agent-to-tool communication — adopted by Salesforce, ServiceNow, SAP, and the major SaaS platforms — it becomes critical infrastructure. If it remains primarily a Microsoft-ecosystem standard, its value is diminished. Track which enterprise platforms ship native MCP servers.

A2A cross-platform reality. Watch for Google Cloud, AWS, and independent framework builders implementing A2A. The protocol's viability as an interoperability standard depends entirely on ecosystem breadth. If only Microsoft agents speak A2A, it is a proprietary protocol with open documentation.

Agent Factory pricing announcement. The commercial model for the integrated platform will define adoption patterns. If pricing is consumption-based and transparent (like SRE Agent), adoption will be broad. If pricing is opaque enterprise licensing with large minimum commitments, adoption will be limited to organisations that have already decided to go all-in on Microsoft.

EU AI Act compliance tooling. As the AI Act enforcement timeline approaches, agent governance frameworks will need to map directly to regulatory requirements. The first vendor to ship compliance-specific governance templates and automated regulatory reporting for AI agents gains a significant advantage. Microsoft's governance infrastructure positions them well here, but execution is everything.

Independent SRE Agent validation. Microsoft claims 20,000 engineering hours saved internally. Independent validation from external deployments — with real incident data, actual autonomous resolution rates, and honest assessments of failure modes — will determine whether those numbers translate outside of Microsoft's own infrastructure.

Open-source framework maturity. LangChain, CrewAI, and the broader open-source agent ecosystem are moving fast. If these frameworks deliver "good enough" multi-agent orchestration with MCP support and without platform lock-in, they become a credible alternative for organisations unwilling to commit to Microsoft's integrated platform. Watch the gap between open-source capabilities and Microsoft's platform benefits — if it narrows, the lock-in premium becomes harder to justify.

Analysis from Microsoft Ignite 2025, San Francisco, 18-21 November. Steven Newall is VP of Engineering with 21 years experience in enterprise infrastructure, platform engineering, and cloud. He attended Ignite 2025 as a practitioner evaluating these technologies for real deployment, not as a Microsoft partner or evangelist.