Microsoft previews Microsoft Agent Framework
Microsoft announced the public preview of Microsoft Agent Framework, an open-source SDK and runtime for multi-agent orchestration.
By Hypler editorial team · Source: Microsoft.
Original source date: .
Hypler publication: . Hypler updated: .
Estimated reading time: 4 minutes.
Topics: Microsoft, agents, orchestration, Azure AI Foundry

The preview described components for composition
Microsoft announced the public preview of Microsoft Agent Framework on October 1, 2025, describing an open-source SDK and runtime for agent and workflow composition. The announcement connected it with Azure AI Foundry observability and governance capabilities and said multi-agent workflows were in private preview. Documentation describes model clients, sessions, context providers, middleware, MCP clients, and workflows. Application teams still choose those components, define their contracts, and evaluate the complete workflow. A framework release does not decide whether a proposed set of agents has a coherent authority model.
Coordination needs one accountable owner
Each agent adds choices about task selection, visible state, delegation, tool use, and disagreement resolution. Assign one workflow owner even when several components participate. That owner defines shared schema, terminal states, retry rules, and the point where a human decides rather than letting coordination continue indefinitely. Specialization is valuable when it creates a measurable boundary: retrieval can return cited material, planning can return a non-executable plan, and validation can check deterministic rules. Calling every component an agent can hide those distinctions. Prefer the simplest component that satisfies the contract and make each handoff inspectable.
A policy-review workflow can stop safely
Imagine a hypothetical policy review. Intake validates file type and classification. Retrieval selects allowed passages and returns source identifiers. Drafting produces a labeled proposal that cannot send, publish, or change a record. A deterministic checker verifies headings, citations, and prohibited terms. A reviewer receives the proposal, sources, and checker result and owns approval. If retrieval cannot establish an allowed source, drafting does not invent a substitute; if the checker fails, the artifact stays incomplete. The important outcome is not the number of components but their narrow inputs, outputs, and safe stopping points.
State and middleware need an explicit contract
Treat workflow state as data rather than accumulated conversation. Include task identifier, input version, permitted sources, stage, owner, allowed transitions, and terminal reason. Persist only what is needed for audit or resumption; do not give every component unrestricted history. A retry retains identity and attempt count so an interruption does not become several independent actions. Middleware can enforce policy around model and tool calls, but tool servers should validate authority themselves. Emit events for stages, tool requests, failures, and human decisions without capturing sensitive prompts by default.
Orchestration can multiply ambiguity
Multi-step work can loop on retries, pass stale context forward, or let one output masquerade as an instruction to another component. Concurrent branches may race on shared state. A tool call can succeed after its coordinator timed out, leaving the visible status wrong. Design for cancellation, maximum attempts, stable ordering, and explicit conflict state. Test failed calls, timeouts, injected instructions in retrieved text, incompatible updates, and human rejection. The trace should show where the workflow stopped, what was preserved, and what a reviewer needs for a safe retry, rather than treating a final model message as authoritative system state.
Release evidence belongs to the whole workflow
Test normal completion, missing evidence, invalid transitions, duplicate delivery, tool denial, and escalation. Inspect whether outputs trace to allowed inputs, state survives restart, and side effects stop after failed validation. Use deterministic tool fixtures and retain versions of models, agents, tools, and policies. A candidate needs owners for schemas, tool rules, logs, updates, and rollback. Begin before a system-of-record boundary, monitor stop reasons and unresolved outcomes, and rerun integration fixtures after framework changes. Hypler's analysis treats observability as useful only when it supplies this operational evidence. This keeps automation subordinate to a checkable workflow.
Hypler analysis
Hypler’s view
More agents add more decision boundaries. The framework becomes useful when a workflow has one accountable owner, explicit handoffs and terminal states, and traces that distinguish a model suggestion from an authorized system action.
Hypler technical analysis
Engineering notes
Multi-agent orchestration introduces state transfers, delegated tool selection, and exception paths that may not be visible in a single prompt. An SDK and runtime should therefore be modeled as a responsibility graph: every handoff needs a bounded input, owner, permitted effects, observability record, and recovery decision.
What to inspect
- Responsibility map
- Inspect which agent may plan, retrieve, select a tool, write state, request approval, or cause an external effect; reject overlaps without a defined owner.
- State handoff
- Trace task state across agents, including identity, data provenance, authorization scope, revision, failure state, and cancellation propagation.
- Exception trace
- Trigger tool failure, conflicting outputs, timeout, and invalid handoff; inspect whether the orchestration preserves evidence and routes the case to a human decision point.
Scope and limits. Preview framework features and enterprise positioning do not prove traceability or governance in a particular multi-agent implementation; those outcomes depend on workflow design and retained evidence.
Terms and definitions
- MCP
- Model Context Protocol, an open protocol for connecting AI applications to tools, data sources, and prompts through defined interfaces.
- Tool Calling
- A model interface pattern where the model requests a structured operation from an external tool.
- API Layer
- The server-side interface that exposes defined data and actions to clients or integrations.
- Testing and Observability
- Testing checks expected behavior; observability uses runtime signals to understand a system’s actual behavior and condition.
- Google announces the Agent Development Kit
Google announced the Agent Development Kit, an open-source framework for building and evaluating multi-agent applications.
- Microsoft previews Deep Research in Azure AI Foundry
Microsoft announced a public preview of Deep Research in Azure AI Foundry Agent Service. The company described an API and SDK offering based on OpenAI's research capability, with Bing Search grounding and source-backed output.
- Anthropic introduces web search on the API
Anthropic introduced web search as an API tool for supported Claude models, returning web results and citations for applications using current public sources.

Original source
Historical source briefing with Hypler engineering analysis. Source details reflect the original publication date; implementation examples are hypothetical. Technical analysis expanded October 24, 2026; this does not change the original source event or establish a new source verification. Editorial prose revised October 26, 2026; original event, publication and source-check dates are retained.