Build 2026 Foundry agents: what Azure teams can ship now

Microsoft Build 2026 opened yesterday at Fort Mason in San Francisco, running 2–3 June, with Satya Nadella's keynote streamed from 9:30 a.m. Pacific time. For teams building on Azure, this was the Build where Microsoft Foundry stopped being a model catalogue with an agent SDK bolted on and became a full production-agent stack: hosted runtimes, a unified tool endpoint called Toolboxes, agent memory, real-time voice, and Foundry IQ, a single SLA-backed retrieval endpoint that unifies Work IQ, Fabric IQ, Azure SQL, File Search and MCP sources. For anyone running retrieval-augmented generation on Azure, Foundry IQ is the consequential one. Microsoft also showed Scout, its first “Autopilot” agent that runs continuously in the background across Microsoft 365. The catch, as always at Build, is that the announcements span everything from generally available to “early experimental release”. If you are planning your 2026 agent architecture from Stockholm or anywhere else in the EU, the status labels matter more than the demos.

What Microsoft actually announced

Foundry IQ: one retrieval endpoint instead of five grounding pipelines

The most architecturally significant announcement is Foundry IQ. Until now, grounding an Azure-hosted agent in enterprise data meant stitching together separate pipelines: Azure AI Search for documents, a Fabric connection for analytics data, custom connectors for SQL, Graph connectors for Microsoft 365 content, and increasingly MCP servers for everything else. Foundry IQ collapses that into one retrieval endpoint that unifies Work IQ (Microsoft 365 work data), Fabric IQ, Azure SQL, File Search and MCP sources, with agentic retrieval on top.

The status split matters. Foundry IQ knowledge bases, the multi-source knowledge bases you configure and query, are generally available. Foundry IQ Serverless, the consumption-based unified retrieval endpoint, is in public preview, as are the newest agentic retrieval improvements and security updates. Microsoft also announced Web IQ, live web grounding with sub-200ms response times and zero data retention, though that is an announcement rather than something you can deploy today.

Why this changes RAG design on Azure: most of the custom retrieval orchestration code Azure teams wrote in 2024–2025 (query routing across sources, per-source ranking, permission trimming) is exactly what Foundry IQ now claims as platform responsibility, behind an SLA. If you are starting a new agent project this quarter, the default question flips from “which retrieval pipeline do we build?” to “which sources do we register with Foundry IQ, and what do we still need to own ourselves?”

Toolboxes: one managed endpoint for every tool type

Toolboxes, now in public preview, address the other sprawl problem: tools. Instead of each agent carrying its own configuration of MCP servers, OpenAPI specs, file search and web search, a Toolbox is a centrally managed endpoint that unifies access to web search, file search, MCP, OpenAPI and A2A protocols, with versioning and governance built in. Notably, Toolboxes include intelligent tool selection to reduce model context overhead: the platform decides which tool definitions to surface to the model rather than stuffing every tool into every prompt. Anyone who has watched an agent's token bill climb as its tool list grew will appreciate why this is a platform feature and not an afterthought.

Agent memory: procedural, user and session

Foundry's agent memory, in public preview, comes in three types: session memory (conversation context), user memory (preferences that persist across sessions), and the new one, procedural memory, which lets agents learn how to do the work across runs, not just remember outcomes. Microsoft cites 7 to 14 percent absolute success-rate gains at near-baseline cost from procedural memory. If that holds up in real workloads, it is a meaningful number: it means a long-running agent gets measurably better at a recurring task without fine-tuning or prompt surgery.

Voice Live: real-time voice agents, generally available

Voice Live prompt agents are generally available. The service bundles speech recognition, text-to-speech, turn detection, interruption handling and avatars into one real-time pipeline. For Swedish organisations that have been prototyping voice assistants by hand-wiring Azure Speech to a chat model, then dealing with the latency and barge-in problems of a DIY pipeline, this is the piece that moves voice agents from demo to deployable.

Hosted agents, Routines and publishing to Teams

On the runtime side, Foundry Agent Service now offers hosted agents with sandboxed sessions, state management and filesystem access, with general availability expected by early July. Routines, in public preview, add scheduled agent execution: the overnight ticket-triage pattern as a platform primitive rather than a Logic App you maintain. The service supports the frameworks teams actually use, including Semantic Kernel, AutoGen and CrewAI. And one-click publishing of Foundry agents into Microsoft Teams and Microsoft 365 Copilot, with identity and permissions applied automatically, is planned for general availability this month, per Microsoft's Foundry blog.

Scout, Autopilot agents and new models

The headline demo was Microsoft Scout, the company's first “Autopilot” agent: an always-on agent that works continuously in the background across Teams, Outlook, OneDrive, SharePoint and local device actions, taking action without being prompted each time. It runs with identity governance through Microsoft Entra. Status is the key detail: Scout is available to Frontier organisations through an early experimental release. That is two qualifiers away from anything you should put in a 2026 plan.

On models, Microsoft put four first-party MAI models into public preview: Thinking 1, Image 2.5, Transcribe 2 and Voice 2. It also made Fireworks AI access generally available and continues to offer Anthropic's Claude Opus through Foundry. Model routing across OpenAI, Anthropic, Mistral and DeepSeek was a running theme of the event. The multi-model Foundry story is now credible enough that hard-coding a single provider into your agent architecture looks like a 2024 decision.

The status board as of 3 June

Stripping away the keynote energy, here is where everything sits as of today, 3 June:

Generally available now: Foundry IQ knowledge bases (multi-source), Voice Live prompt agents, Fireworks AI model access. Agent 365, Microsoft's enterprise control plane for agents, already reached GA on 1 May.
GA committed, weeks away: one-click agent publishing to Teams and Microsoft 365 Copilot (planned June); hosted agents in Foundry Agent Service (expected early July).
Public preview: Toolboxes, agent memory (including procedural memory), Foundry IQ Serverless, Routines, the four MAI models.
Announced or experimental: Web IQ (announced); Microsoft Scout (early experimental release, Frontier organisations only).

A decision framework for your 2026 agent architecture

We suggest sorting every Build announcement into one of three lanes before it touches your roadmap. For each lane, the test is what happens to your production system if Microsoft changes the API, the price or the availability. How impressive the demo was has no bearing on that.

Lane 1: Build on it now (GA, SLA-backed)

Foundry IQ knowledge bases for new grounding work. If you were about to write another custom multi-source retrieval orchestrator, stop and evaluate this first.
Voice Live for any voice-agent initiative. The DIY speech-to-model pipeline is now the option that needs justifying.
Agent 365 as the governance layer. It has been GA for a month and answers the “who audits the agents?” question your CISO is already asking.

Lane 2: Pilot behind an interface (public preview)

Toolboxes, agent memory, Routines and Foundry IQ Serverless are all worth piloting this quarter, but behind your own abstraction, so a preview-period breaking change is a one-file fix rather than a rewrite. Something as simple as this is enough:

// Illustrative: isolate preview services behind your own contract
interface RetrievalProvider {
  ground(query: string, opts: GroundingOptions): Promise<GroundedResult>;
}

class FoundryIqProvider implements RetrievalProvider { /* preview API here */ }
class LegacyAiSearchProvider implements RetrievalProvider { /* current pipeline */ }

// Agents depend on RetrievalProvider, never on the Foundry SDK directly.
// Swapping providers - or rolling back - is a DI configuration change.

Procedural memory deserves a specific pilot design: pick one recurring, measurable task (triage, classification, a weekly report), run it with and without memory enabled, and check whether you see anything like the claimed 7–14 point success-rate improvement on your workload before you plan around it.

Lane 3: Watch, don't build (announced/experimental)

Scout and the Autopilot category, and Web IQ. An always-on agent that acts across Outlook, SharePoint and local devices without per-action prompts is precisely the class of system where you want Microsoft and the early experimental cohort to find the failure modes first. Track it, brief your security team on the concept, and revisit when there is a public preview with EU-relevant terms.

One-page checklist before any Build 2026 feature enters your architecture: (1) Is it GA with an SLA, or preview? (2) Is it available in the Azure regions your data must stay in? (3) Does it appear in your enterprise agreement or does preview usage sit outside support terms? (4) Can you wrap it behind an interface you own? (5) Do you have an exit path if the GA pricing surprises you? Five yes/no answers. If you cannot answer all five, treat the feature as a demo you liked, not as architecture.

The Swedish and EU angle

Most Swedish enterprises building agents will consume exactly this stack, because the Microsoft estate (Azure, Microsoft 365, Fabric) is where their data already lives. That makes three EU-specific questions urgent rather than theoretical.

Data residency and the unified endpoint. Foundry IQ's whole value is that one endpoint reaches into Work IQ, Fabric, SQL and file stores. That is also its compliance surface: a single retrieval call may now traverse data sources with different residency guarantees. Before adopting it for regulated workloads, get explicit answers from Microsoft on which regions the knowledge-base and serverless retrieval services run in, and where indexes and caches are stored. The Build materials do not spell this out, and “EU Data Boundary” coverage for brand-new services is something to verify, not assume. The same applies doubly to Web IQ's zero-data-retention claim: promising, but ask for it in writing for your tenant.

The AI Act likes this direction. The general-purpose AI obligations have applied since August 2025, and organisation-level accountability for AI systems is the direction of travel. A platform stack where agents carry Entra identities, run under Agent 365 governance, and publish to Teams with permissions applied automatically is materially easier to document and audit than a fleet of bespoke agents with shared service principals. When your DPO asks “which agents exist, what can they access, and who approved them?”, you want the answer to be a portal view, not a spreadsheet.

Procurement and preview terms. Public preview services typically carry no SLA and different support terms. For Swedish public-sector organisations in particular, that distinction decides whether a service can be used with real data at all. Our practical rule: previews run against synthetic or non-sensitive data only, and the upgrade to production data happens at GA, after residency and terms are confirmed, not when the demo works.

Conclusion

Build 2026's Foundry story is coherent in a way previous years' agent announcements were not: retrieval (Foundry IQ), tools (Toolboxes), memory, voice (Voice Live), runtime (hosted agents, Routines), distribution (Teams/M365 publishing) and governance (Agent 365, Entra) now each have a named platform answer. For an Azure-first EU team, the sensible 2026 posture is to build new grounding and voice work on the GA pieces immediately, pilot Toolboxes and memory behind interfaces you own, and let someone else be the experimental cohort for Autopilot agents. The advantage goes to the team that could tell, on 3 June, which announcements were architecture and which were theatre.