Microsoft Ignite 2026 Preview: Azure AI Foundry, Agents, Copilot and What Platform Teams Should Watch (Nov 17-20)
Most of what platform teams will be asked to evaluate after Microsoft Ignite 2026 has already shipped. In the eight weeks before the conference, Microsoft moved agent toolboxes, tool search, A2A and routines in Foundry Agent Service to general availability, put a GitHub-built agent harness into public preview with a published sandbox price, split Microsoft 365 Copilot into seat and usage billing, and pushed a Kubernetes 1.37 LTS release through AKS. Ignite, which runs November 17-20 at Moscone Center in San Francisco, will add to that pile, but the architectural decisions are already on the table.
That matters because Ignite weeks generate pressure to “do something with agents” before anyone has separated GA from preview from rumour. This preview does that separation with dated, sourced facts, then gives you a launch-evaluation framework, an adopt/pilot/wait table, a worked cost model and a 90-day plan.
What this covers: the verified pre-Ignite baseline, the Microsoft agent stack as it stands, what to watch at the show, cost and lock-in mechanics, and a practical plan.
Context and Background
Ignite is Microsoft’s annual conference for IT pros, platform engineers and developers. The official Ignite site lists the event as November 17-20, 2026 in San Francisco, and the Moscone Center confirms it spans Moscone North, South and West with a digital option. The opening keynote session page (KEY01) lists Satya Nadella and Judson Althoff as speakers. Third-party coverage reports it at 10:00 am PT on Tuesday, November 17, at Chase Center, but Microsoft notes session times are subject to change, so treat the slot as reported rather than confirmed.
Naming first, because it trips up search and procurement alike. What most people still call Azure AI Foundry was renamed Microsoft Foundry at Ignite 2025, and Microsoft Learn now uses “Microsoft Foundry” throughout, with “Foundry Agent Service” as the agent runtime inside it. URLs still carry /azure/foundry/, and older contracts and Terraform modules still say Azure AI Foundry. This post uses “Foundry” for the platform and “Azure AI Foundry” only where the legacy name matters.
The competitive frame is clear. AWS re:Invent runs two weeks later, and our AWS re:Invent 2026 preview covers what Amazon shipped ahead of its show. KubeCon North America landed just before both, and the KubeCon North America 2026 recap documents where the open-source inference and agent-runtime stack actually is. Read together, the three events tell you whether a capability is a hyperscaler differentiator or a commodity you can get from upstream.
This preview takes a narrower angle than either of those posts. It is not a session guide and not a news recap. It treats the Microsoft agent stack as a platform-engineering problem: which layers you will own, which meters will tick, and which announcements change a decision you have to make in the next quarter.
One more ground rule. Everything labelled “verified” below links to or names a Microsoft primary source dated before November 15, 2026. Everything labelled “expected” or “analysis” is my reading of the session catalogue and pre-show signals, and none of it is an announcement. The Ignite Book of News will settle those questions on day one.
The Verified Pre-Ignite Baseline: Microsoft’s Agent Stack as of November 2026
Direct answer: Going into Microsoft Ignite 2026, Microsoft’s agent platform has four GA layers already in production use: Foundry Agent Service with hosted agents and toolboxes, Microsoft Agent Framework 1.0 for code-first agents, Agent 365 with Entra Agent ID for governance, and AKS for self-run inference. The newest pieces, the Copilot harness and the AI Gateway tier, remain in preview.

Figure 1: The Microsoft agent stack going into Ignite 2026, from experience layer to infrastructure, with the control plane cutting across.
The diagram reads top to bottom. Users reach agents through Microsoft 365 Copilot, Teams or Copilot Studio. Those agents execute in Foundry Agent Service, either as declarative prompt agents, as containerised hosted agents, or on the new Copilot harness. Code-first teams build with Microsoft Agent Framework and deploy into the same service. Model calls go through Foundry Models and Model Router, ideally fronted by the API Management AI Gateway tier, while tools arrive through Toolboxes, MCP (Model Context Protocol) and A2A (Agent-to-Agent). Entra Agent ID and Agent 365 apply identity and policy, and AKS or Azure Local hosts anything you run yourself.
Foundry Agent Service moved from preview to production in 2026
The July 9, 2026 Azure blog post announcing GPT-5.6 in Foundry also marked a turning point for the agent runtime. It declared hosted agents, hosted agents with Voice Live, Foundry IQ, toolboxes, and tracing and evaluation for hosted agents generally available. Publishing agents to Teams and Microsoft 365 Copilot followed a week later. Memory, routines and the agent optimizer were public preview at that point.
The September 24 Foundry announcement pushed further. Toolboxes became GA for hosted agents and entered public preview for prompt agents. Tool search, A2A and routines went GA. Voice agents and long-running resilience entered public preview, and Insights, which clusters trace behaviour into findings, entered public preview. Rubric evaluation, synthetic dataset generation and the agent optimizer were targeted for GA “later this month”, and Microsoft’s own October 15 recap still treated those as targets rather than confirmed releases.
October added two runtime pieces. On October 19, the browser automation tool in Foundry became GA. On October 28, Microsoft introduced the Copilot harness in Foundry Agent Service, discussed below, and stated that long-running resilience for hosted agents is now generally available.
The practical meaning: the “is Foundry Agent Service production-ready” debate is over for hosted agents. The debate for Ignite is about the newer, more opinionated runtime and the commercial model around it.
Microsoft Agent Framework is the code-first path
Microsoft Agent Framework 1.0 reached general availability in early April 2026 (Microsoft’s announcement is dated April 3). It merges Semantic Kernel and AutoGen into one .NET and Python SDK. Version 1.0 shipped stable agent abstractions, middleware, memory and context providers, graph-based workflows with checkpointing and human-in-the-loop steps, and orchestration patterns including sequential, concurrent, handoff, group chat and Magentic-One.
MCP support was in 1.0. Coverage at launch said A2A 1.0 support was “coming soon”, and A2A was GA in Foundry Agent Service by September. AutoGen is now in maintenance mode, receiving bug fixes and security patches, while Semantic Kernel remains active with migration guides toward the new framework.
By August, InfoQ reported that the Agent Harness inside the framework and Foundry hosted agents had both reached GA. The September Foundry recap lists Python AG-UI support as stable and .NET AG-UI as preview, at Python 1.19.0 and .NET 1.22.0, with some connectors still alpha or preview. That connector maturity gap is a real adoption detail, not a footnote.
The Copilot harness is the biggest unknown going into Ignite
The Copilot harness in Foundry Agent Service is an optional managed runtime for prompt agents. Microsoft runs the agent loop and a sandbox, and you supply instructions, skills, tools and a supported Foundry model. Built-in capabilities include file read and write, shell, web access and subagents, and tools can be Foundry tools or MCP servers with explicitly configured authentication and approvals.
It is public preview with no SLA. Runtime use is free until December 1, 2026. From that date, Microsoft lists sandbox compute at $0.0994 per vCPU-hour and memory at $0.0118 per GiB-hour, with model inference and tools billed separately, and rates may vary by region and agreement.
Preview limits matter for design. The harness must be chosen at agent creation and existing agents cannot switch. Foundry manages sandbox CPU, memory and idle settings, so you cannot size them. The sandbox is not durable storage. Microsoft also says the harness supports 20 or more frontier models and bring-your-own-key, while third-party coverage of Microsoft Learn says bring-your-own-model is unsupported in the managed preview. Check the docs for your model before you plan around it.
Microsoft reports internal evaluation gains over its standard harness of 10.8 percentage points with an OpenAI model and 22.1 points with an Anthropic model. Those are vendor-reported and not independently verified. Treat them as a reason to run your own evaluation, not as a result.
Governance, gateway and identity are converging
Agent 365 reached GA on May 1, 2026 at $15 per licensed user per month, with Entra Agent ID giving each agent its own identity for lifecycle and access management. The September Foundry release confirmed Foundry agents support governance operations from Entra and Agent 365, without a separate maturity label.
The API Management AI Gateway tier entered public preview on July 27, 2026, in East US 2 and Sweden Central, free during preview. It publishes models from Foundry, Bedrock, Vertex and direct providers behind one endpoint, exposes or federates MCP servers, and applies token rate limits, quotas, content safety and model fallback as policy cards. It emits OpenTelemetry token metrics. That last point lines up with the OpenTelemetry GenAI semantic conventions most observability stacks now speak.
The Rest of the Pre-Ignite Baseline: Copilot, Fabric, AKS and Sovereign
Agents get the headlines, but four other workstreams moved in the run-up to Microsoft Ignite 2026. Each changes a platform decision.
Copilot pricing split seats from usage
Message center post MC1479276, published September 25, 2026, and Microsoft’s Copilot pricing model post describe a split. “Everyday AI” stays inside the per-user Copilot licence. “Advanced AI”, including new frontier models, may be billed through usage-based billing (UBB) in Copilot Credits.
The control point is the spending policy. In Enterprise tenants, UBB stays inactive until an admin creates a spending policy in the Microsoft 365 admin center, and only then are credits consumed. Rollout begins in Q4 CY2026. Spending policies can assign models and effort levels by group, cost management extends to Copilot Studio agents from October, and the listed FinOps capabilities are preview.
For platform teams, this creates a second AI meter that finance will ask you to reconcile with the Azure bill. Foundry agents bill tokens and compute to an Azure subscription. Copilot agents bill credits to a Microsoft 365 tenant. The same business process can land on either meter depending on where the agent was built.
GitHub Copilot SDK becomes Microsoft Copilot SDK
The GitHub changelog of October 28, 2026 states that the GitHub Copilot SDK is becoming the Microsoft Copilot SDK, with no immediate changes needed for existing users. The same day, the Copilot Studio harness was renamed from GitHub Copilot harness to Microsoft Copilot harness, with existing agents unchanged in capability and billing. The pattern is consolidation: one harness, one SDK name, surfaced in Copilot Studio, Foundry and GitHub. Package-name and namespace cut-over dates had not been published as of mid-November.
Fabric moved its agent story at FabCon Europe
FabCon and SQLCon in Barcelona ran September 28 to October 1, 2026. Reported GA items include Fabric IQ business context in Copilot Chat, operations agents root-cause analysis, Google BigQuery mirroring, and DiskANN vector indexes across Hyperscale, Managed Instance and SQL database in Fabric. Preview items include a Fabric data engineering agent, observability in Fabric, Database Hub and SQL and PostgreSQL database agents. Foundry’s own docs list “Connect agents to Microsoft Fabric with Fabric IQ” as preview.
The architectural signal is that grounding data for agents is moving into the data platform as a semantic layer, not into a separate vector store. If you are designing that layer yourself, our analysis of semantic layers for LLM analytics covers the governed-metrics side of the same problem.
AKS shipped Kubernetes 1.37 LTS and a NAT change
The AKS release of October 25, 2026 made Kubernetes 1.37 GA and designated it LTS (long-term support), with 1.36 as the new-cluster default. Node Auto Provisioning gained a batch of GA features, including capacity reservation groups, a capacity buffers CRD, managed GPU and Azure Container Linux through AKSNodeClass. Kubernetes 1.34 community support and 1.31 LTS support end by November 2026, and from June 1, 2027 AKS will auto-upgrade clusters left on unsupported versions after the 60-day grace period.
Separately, Azure update AZ574430 (October 8) made a managed StandardV2 NAT Gateway the default for new AKS clusters using the managedNATGateway outbound type, with zone redundancy and IPv6. For the cross-cloud comparison, see our EKS vs GKE vs AKS breakdown.
Sovereign and edge: Foundry Local on Azure Local
An Azure Arc blog post dated October 20, 2026 says Foundry Local is “entering general availability on Azure Local at Ignite”, alongside GitHub Enterprise Local. The model catalog and agentic retrieval are listed as GA, while agents with enterprise connectors and configurable guardrails remain preview. This is one of the few items where Microsoft has publicly pre-announced an Ignite GA, so expect it, but verify the Book of News wording and the region and SKU list. Our data residency and sovereign cloud architecture post covers the design patterns this enables.
Azure pricing: a February 2027 regional increase
Customers were notified through Azure Service Health in early October of price increases effective February 1, 2027 in seven regions, reported as 9 percent in West Europe, 17 percent in North Europe, 11 percent in the two France regions, 9 percent in the two Norway regions and 8 percent in Southeast Asia. Microsoft published exemptions, reportedly including v7 VMs, rather than a full list of affected services. A reported 25 percent global rise for v1 and v2 series comes from a partner, not Microsoft. Treat the details as reported until your own notices confirm them.
What to Watch at Ignite 2026: A Launch-Evaluation Framework
The volume of Ignite 2026 announcements will exceed any team’s capacity to evaluate them. The solution is a filter that runs in minutes per announcement and produces a defensible decision.

Figure 2: A triage flow for Ignite 2026 announcements, keyed on status label, SLA and region fit, exit cost and evaluation results.
The flow starts with the status label, because it determines everything else. A GA feature still needs an SLA check and region fit, then an exit-cost test. A preview feature goes to a sandbox with no production data and must clear an evaluation bar within 30 days. Anything only “expected” goes on a watchlist and gets re-scored once the Book of News is out.
Score each announcement on five questions
Run every announcement through five questions and write the answers down.
- Status. GA, public preview, private preview, or “coming soon”? Microsoft’s own blogs sometimes give no maturity label, as the September post did for Agent Framework updates and Entra lifecycle operations. No label means preview until proven otherwise.
- SLA and support. GA services carry an SLA, preview features usually do not, and the Copilot harness preview explicitly has none.
- Region fit. The harness lists nine regions. The AI Gateway tier preview lists two. Model Router covers 32. A feature unavailable in your data-residency region is not available to you.
- Meter. Which bill does it land on: Azure consumption, Copilot Credits, a per-user licence such as Agent 365, or a free preview that will start charging? The harness free period ends December 1, 2026, two weeks after Ignite.
- Exit cost. If you build on this and leave in 18 months, what do you rewrite? Agent definitions in a Microsoft-specific declarative format are expensive to move. Code written against Agent Framework with MCP tools is cheaper.
What I expect Ignite to address (analysis, not announcement)
The following are expectations drawn from the session catalogue and the pre-show trajectory. None of them has been announced.
- Harness maturity. The Copilot harness launched three weeks before Ignite with a December 1 billing start. A roadmap to GA, more regions or switchable harness selection would be the natural next step.
- One agent inventory. Agent 365 governs Microsoft 365 agents and Foundry agents support its lifecycle operations. A single inventory across both is the obvious gap.
- AI Gateway tier progression. The tier launched in two regions with pricing “to be shared separately”, and a Foundry portal experience was promised.
- Evaluation and optimizer GA confirmation. Rubric evaluation, dataset generation and the agent optimizer were targeted for September GA without clear confirmation.
- Copilot SDK cut-over dates. Package and namespace changes are pending.
- Foundry Local on Azure Local GA, which Microsoft has pre-announced.
If an item on this list lands, run it through the five questions. If it does not, nothing in your current plan should depend on it.
Deeper Analysis: Cost Mechanics of an Agent Call
The least understood part of the Microsoft agent stack going into Microsoft Ignite 2026 is how many meters a single agent task touches. Figure 3 traces one task.

Figure 3: The billing meters that tick during one agent task, from a Teams request through the gateway, model, harness sandbox and MCP tool.
A user in Teams asks an agent to do something. If the agent lives in the Microsoft 365 Copilot path, Copilot Credits may apply under a spending policy. The agent calls a model through the AI Gateway, where tokens in and out are metered. If it runs on the Copilot harness, sandbox vCPU and memory meters run for as long as the sandbox is active. Each MCP tool call may hit a connector or downstream API with its own price. The trace ties it together, but only if you propagate a session or trace ID end to end.
A worked cost model (illustrative)
The following numbers are illustrative. They use Microsoft’s published list rates as of November 2026 and assumed workload sizes. Your sandbox sizing is not controllable in preview, so the CPU and memory figures are assumptions, not observed values.
Assume a document-triage agent on the Copilot harness running GPT-6.1 Sol, which Microsoft’s September recap lists at $2.00 per million input tokens, $0.10 per million cached input tokens and $10.00 per million output tokens. Assume each task uses 60,000 input tokens, 8,000 output tokens and a sandbox of 2 vCPU and 4 GiB active for 10 minutes. Run 1,000 tasks per day.
| Cost component | Calculation | Per task | Per day (1,000 tasks) |
|---|---|---|---|
| Sandbox compute | 2 vCPU x $0.0994 x 1/6 h | $0.0331 | $33.13 |
| Sandbox memory | 4 GiB x $0.0118 x 1/6 h | $0.0079 | $7.87 |
| Input tokens, no caching | 60,000 x $2.00 / 1M | $0.1200 | $120.00 |
| Output tokens | 8,000 x $10.00 / 1M | $0.0800 | $80.00 |
| Total, no caching | $0.2410 | $241.00 | |
| Input tokens, 50% cached | 30,000 x $2.00 / 1M + 30,000 x $0.10 / 1M | $0.0630 | $63.00 |
| Total, 50% cache hits | $0.1840 | $184.00 |
Three things fall out of this. First, the sandbox is about 17 percent of the uncached total, so the December 1 billing start matters but model tokens dominate. Second, prompt caching is the biggest single lever, cutting about 24 percent of the total here. Third, sandbox idle time is a hidden multiplier. If the managed sandbox stays alive 30 minutes rather than 10, compute and memory triple to roughly $123 per day, and you cannot tune idle settings during the preview.
Why session affinity is a cost feature
Model Router’s preview session affinity, described in Microsoft’s October 15 recap, prefers the same eligible model across related requests to help prompt-cache reuse. Microsoft says neither a cache hit nor the same model is guaranteed. The design implication is concrete. Reuse one session ID for all turns of a task, use separate IDs for unrelated work, and log the routing metadata so you can see when a fallback broke the cache. In the model above, losing cache reuse costs about $57 per day at 1,000 tasks.
Two meters, one business process
The harder problem is reconciliation. If marketing builds an agent in Copilot Studio and engineering builds the same workflow in Foundry, one consumes Copilot Credits under a Microsoft 365 spending policy and the other consumes Azure tokens and compute. They will show up in different admin centres, cost reports and budget owners.
A pragmatic pattern is to tag every agent with an owning cost centre in Agent 365 or Entra Agent ID metadata, export both meters into the same FinOps store, and join on agent identity. The cloud cost anomaly detection patterns apply directly once both feeds are in one place. Without the join, nobody can answer “what does this agent cost per task?”
Decision Matrix: Adopt, Pilot or Wait
Ahead of Microsoft Ignite 2026, the table below applies the five-question framework to what is verifiable today. It will need updating after the Book of News, which is the point: the decision record should change when the facts change, not when the keynote does.
| Capability | Status (verified, pre-Ignite) | Main risk | Recommendation |
|---|---|---|---|
| Foundry Agent Service hosted agents | GA, long-running resilience GA | Container and identity sprawl | Adopt for new agent workloads, behind a gateway |
| Toolboxes, tool search, A2A, routines | GA (prompt-agent toolboxes preview) | Tool permission creep | Adopt with per-tool approval policy |
| Microsoft Agent Framework 1.0 | GA, some connectors alpha or preview | Connector maturity varies | Adopt for code-first agents; check each connector |
| Copilot harness | Public preview, no SLA, billing from Dec 1 | Unsizable sandbox, no switching, shell reach | Pilot in sandbox subscriptions only |
| APIM AI Gateway tier | Public preview, two regions, price TBD | Pricing unknown, region limits | Pilot; use classic APIM AI gateway policies in production |
| Model Router session affinity | Preview | Silent fallback breaks cache | Pilot with routing metadata logged |
| Agent 365 and Entra Agent ID | Agent 365 GA since May 1 | Licence model is per human user | Adopt if you run more than a handful of M365 agents |
| Copilot usage-based billing | Rolling out Q4 CY2026, FinOps preview | Second AI meter | Adopt controls now: spending policies before enablement |
| Fabric IQ for agents | Copilot Chat business context GA, Foundry connection preview | Semantic model quality | Pilot where Fabric is already the system of record |
| AKS 1.37 LTS | GA | Upgrade cadence, Cilium Kafka policy removal | Adopt for long-lived inference clusters |
| Foundry Local on Azure Local | Pre-announced GA at Ignite | Wording and SKU list unknown | Wait for Book of News, then pilot |
“Adopt” here means approved for new production workloads with standard controls, not “migrate everything”. “Pilot” means a time-boxed evaluation with a written exit criterion. “Wait” means no engineering spend until a named event happens.
Lock-in: Where the Real Switching Costs Sit
Lock-in is the question Microsoft Ignite 2026 keynotes will not answer for you. Lock-in in the agent era is not where it was for VMs or even Kubernetes. Models are increasingly swappable, since Foundry hosts OpenAI, Anthropic, Mistral and others. The switching cost has moved up the stack into four places.
Agent definitions. Declarative agent definitions in Foundry or Copilot Studio encode instructions, tool bindings, approval policies and harness choice in a platform-specific format. Microsoft says harness agent definitions can be kept in source control, which helps auditing but does not make them portable. A Microsoft Agent Framework codebase using MCP tools is the most portable option Microsoft offers, because the framework also ships connectors to Bedrock, Gemini, Anthropic and Ollama.
Tool and connector surface. The AI Gateway tier exposes more than 1,400 connector-backed tools from Power Platform and Logic Apps. That is convenient and sticky. An MCP server you own can be called from any MCP-capable runtime. A Logic Apps connector wrapped as an MCP tool still depends on Logic Apps.
Identity and governance. Entra Agent ID and Agent 365 give each agent an identity, a lifecycle and an inventory. Once your audit and access reviews depend on that inventory, moving agents elsewhere means rebuilding governance. That is fine if Entra is already your identity plane, which it is for most Microsoft shops. For the general pattern, see our post on agent identity and delegated authorization.
Evaluation and trace data. Insights, rubric evaluators and the agent optimizer improve agents based on Foundry traces. If traces only live in Foundry, the improvement loop is locked in. Exporting traces in OpenTelemetry form to a store you own keeps that asset portable.
The mitigation is the same in each case: put a thin layer you own at each seam. A gateway in front of models, MCP servers you control in front of systems of record, OpenTelemetry export for traces, and agent code in a framework with multi-provider connectors. Our LLM gateway comparison covers the gateway options if you would rather not standardise on APIM.
Trade-offs, Gotchas, and What Goes Wrong
Pre-conference previews of Microsoft Ignite 2026 tend to skip failure modes. These are the ones I would plan for.
Shell reach in the harness. Third-party analysis of Microsoft Learn notes that the harness shell can read, write and reach network resources even when the file and web toggles are off. If that is accurate, disabling the file and web tools is not a network control. Apply egress restrictions at the network layer and treat the shell as untrusted code execution. Our agent sandboxing comparison explains why isolation has to sit below the tool layer.
Non-durable sandbox state. The harness sandbox is not durable storage. An agent that writes intermediate results to its working directory will lose them on recycle. Persist outputs to Foundry Files or blob storage at every step that would be expensive to redo.
Preview-to-GA drift. Preview features change shape at GA. The harness cannot switch on existing agents during preview, so anything you build now may need recreating. Budget for that rework or keep preview agents disposable.
Unlabelled maturity. Microsoft’s announcement posts sometimes list features with no GA or preview label, and “later this month” targets do not always land on time. Track status from Microsoft Learn and Azure Updates, not from keynote slides.
Billing start dates near Ignite. The harness starts billing December 1, and Copilot UBB rolls out in Q4. A pilot launched in Ignite week can start costing money before the evaluation finishes. Put budget alerts on pilot subscriptions before the first run.
Spending policy as an enablement switch. In Enterprise tenants, creating a Copilot spending policy is what activates credit consumption. A well-meaning admin who creates one to “set a limit” also turns metering on. Write the policy deliberately and document who owns it.
Kubernetes upgrade pressure. With 1.34 community and 1.31 LTS support ending by November 2026, and auto-upgrade of unsupported clusters from June 1, 2027, clusters running inference or agent workloads need an upgrade plan. Cilium 1.20 no longer enforcing Kafka-aware network policies is the kind of change that breaks silently.
Regional price changes. If your agent workloads sit in North Europe or West Europe, the reported February 2027 increase lands mid-budget-year. Confirm the exemptions against your VM series before reserving capacity.

Figure 4: A 90-day post-Ignite plan in three phases, triage, pilot and decide, each ending in a concrete artefact.
Practical Recommendations
The goal for the 90 days after Ignite is not to adopt the most features. It is to finish the quarter with a decision record for each relevant announcement, a cost-per-task number for every production agent, and no preview feature in a production path without a written exception.
Days 1-30: triage. Classify every Ignite 2026 announcement that touches your estate using the five questions. Build an inventory of existing agents across Copilot Studio, Foundry and any self-hosted frameworks, with owner, meter and monthly spend. Set a rule that preview features stay out of production paths.
Days 31-60: pilot. Put a gateway in front of every model call first, with token budgets per team. Choose at most two pilots, for example one hosted agent on Agent Framework and one harness agent, each with a fixed evaluation set and a cost ceiling. Give every pilot agent an Entra Agent ID and explicit tool approvals.
Days 61-90: decide. Compute cost per task from real traces, not list-price estimates. Write an adopt/pilot/wait decision record for each capability, with the exit plan and the event that would change the decision. Produce runbooks for the agents you adopt.
Checklist for Ignite week:
- [ ] Download the Book of News and tag each item GA, preview or announced-only.
- [ ] Re-score the decision matrix above against the new facts.
- [ ] Put budget alerts on any subscription where the harness is enabled before December 1.
- [ ] Confirm whether a Copilot spending policy exists in your tenant and who owns it.
- [ ] Check AKS cluster versions against the November 2026 support cut-offs.
- [ ] Verify region availability for every feature you plan to pilot.
- [ ] Export Foundry traces to an OpenTelemetry store you control.
Frequently Asked Questions
When and where is Microsoft Ignite 2026?
Microsoft Ignite 2026 runs November 17-20, 2026 at the Moscone Center in San Francisco, with a digital attendance option. Coverage citing the official site lists a pre-day on November 16. The opening keynote session lists Satya Nadella and Judson Althoff as speakers, and third-party coverage reports it at 10:00 am PT on November 17 at Chase Center, though Microsoft says times can change. Most sessions are expected on demand 24 to 48 hours after they end, so remote platform teams can follow the technical tracks without travelling.
Is Azure AI Foundry the same as Microsoft Foundry?
Yes. Microsoft renamed Azure AI Foundry to Microsoft Foundry at Ignite 2025, and current Microsoft Learn documentation uses the new name throughout. Foundry Agent Service is the agent runtime within it. Documentation URLs still include /azure/foundry/, and many contracts, Terraform modules and internal wikis still say Azure AI Foundry, so search both names when you research features. The underlying Azure resources, regions and billing continue under your Azure subscription, so the rename does not change procurement mechanics or existing deployments.
What is the Microsoft Agent Framework and should I migrate from Semantic Kernel?
Microsoft Agent Framework is an open-source .NET and Python SDK that merges Semantic Kernel and AutoGen. Version 1.0 reached GA in early April 2026 with stable agent abstractions, graph workflows, MCP support and multi-agent orchestration patterns. AutoGen is now maintenance-only, so AutoGen users should plan a migration. Semantic Kernel remains active, and Microsoft provides migration guides, so Semantic Kernel teams can migrate on their own schedule. New agent projects should start on Agent Framework, checking connector maturity first.
What will the Copilot harness in Foundry cost?
Harness runtime use is free until December 1, 2026. From then, Microsoft lists sandbox compute at $0.0994 per vCPU-hour and memory at $0.0118 per GiB-hour, with model inference and tools billed separately, and rates may vary by region and agreement. In an illustrative model of 1,000 ten-minute tasks per day on 2 vCPU and 4 GiB, the sandbox costs about $41 per day while model tokens cost about $200. Sandbox sizing and idle settings are managed by Foundry during preview, so actual costs depend on behaviour you cannot tune.
Which Ignite 2026 announcements should platform teams act on first?
Act first on anything that changes a meter or a support deadline. That means Copilot usage-based billing controls, the December 1 harness billing start, AKS version cut-offs and any regional pricing changes. Next, look at GA capabilities that reduce custom code, such as Foundry toolboxes, A2A and routines. Preview features like the harness and the AI Gateway tier deserve a time-boxed pilot, not production commitments. Anything only expected or pre-announced should wait for the Book of News and Microsoft Learn documentation before anyone spends engineering time.
How does Microsoft’s agent stack compare with AWS ahead of re:Invent?
Going into Microsoft Ignite 2026 and re:Invent, both vendors offer a managed agent runtime, a code-first SDK, a gateway for models and tools, and identity for agents, and both support MCP. Microsoft’s distinctive assets are its Microsoft 365 and Teams distribution, Entra-based agent identity and Fabric as a grounding layer. AWS’s are covered in our re:Invent preview, which comes two weeks later. The practical advice is the same for both: keep agent code in portable frameworks, own your MCP servers and traces, and compare cost per task rather than headline features.
Further Reading
- KubeCon North America 2026 recap: AI inference, agentic workloads and what actually shipped for the open-source runtime baseline.
- AWS re:Invent 2026 preview: pre-Invent launches and what platform teams should watch for the competing hyperscaler view.
- LLM gateway comparison: LiteLLM vs Portkey vs Kong AI Gateway vs Cloudflare AI Gateway for gateway options beyond APIM.
- Microsoft Foundry Blog: Copilot harness in Foundry Agent Service, the primary source for harness status and pricing.
- Microsoft Ignite official site for the session catalogue and the Book of News once published.
By Riju — about
