Microsoft Agent Framework Advances to Production-Ready Runtime

Microsoft announced at Build 2026 the general availability of its Agent Harness and Foundry Hosted Agents, a production-ready runtime for AI agents. The company emphasizes that the harness infrastructure, not the model, is the core product, with features like permissions, context management, sandboxing, and telemetry built in. The framework supports local, container, and hosted deployments from a single binary, and integrates with GitHub Copilot and the Claude Agent SDK.

The AI agent story is mostly not an AI story. That's the uncomfortable, useful takeaway from Build 2026, where Microsoft put its Agent Harness and Foundry Hosted Agents into General Availability — a real production runtime, not another experimental SDK. Per recent research on Claude Code v2.1.88, roughly 98.4% of an agent system is harness infrastructure: permissions, context management, sandboxing, tool routing, recovery. Only 1.6% is the model making decisions. Microsoft just shipped that 98.4% as a supported product. That's the announcement worth understanding. Every team that's tried to put a coding agent into production has run into the same bottleneck: the model is the easy part. Wiring it to your tools, persisting conversation history, deciding which actions need human approval, compacting context before you blow the token budget, emitting telemetry you can actually alert on — that's the work. And it was work you were doing yourself, in a bespoke loop, with no support contract and no upgrade path. Microsoft's bet — and it's a correct one — is that the harness is the product. The model is a swap-out part. Two things, announced at Build 2026: Both run locally, in containers, or hosted. One binary, three deployment shapes. The framework also ships stable multi-agent orchestration patterns and connectors for GitHub Copilot and the Claude Agent SDK, both wired into the same governance policy plane. DIAGRAM: a single user message enters the harness, the harness calls the model, tool calls route through an approval gate, history is persisted, context is compacted before the next turn, OpenTelemetry spans emit on every step, the loop returns to the model for the next decision The feature list reads like a checklist of every failure mode that hit teams running agent loops in 2024 and 2025: All of this is enabled by default. Each piece is individually removable. Opt-in features — shell tooling, file access, background sub-agents, automatic looping — still emit warnings when activated. That's the governance posture baked into the binary. The hosted variant exists because most platform teams don't want to run another service. Foundry Hosted Agents is consumption-billed managed deployment: you ship the agent definition, the platform handles the rest. The integration story is the part worth pausing on. You supply: The harness handles everything else through a single API call. Planning, history persistence, compaction, approvals, web search, telemetry — all of it routed through the same code path that runs locally and in containers. There isn't a separate "hosted" mode to learn. This is the shape of a real production story. Same binary in dev, same binary in staging, same binary in prod. The bill changes, the behavior doesn't. The framework ships connectors for both GitHub Copilot and the Claude Agent SDK, and both plug into the same governance policy plane. For coding-agent work specifically, that means a single set of approval rules and audit trails regardless of which model the agent is reasoning through. The connector is the delegation boundary: Agent Framework orchestration calls into the connector, the connector talks to Copilot or Claude, and the result comes back through the same telemetry pipeline. You get observability across model swaps without rewriting your harness. Per the Build 2026 announcement, this is the part Microsoft emphasized for "enhanced governance for coding agents" — the model underneath can change, the audit log doesn't. The release is GA, so the path from zero to a running agent is short. Per the announcement, the Agent Harness is distributed as a single binary that supports local development, containers, and hosted deployments from the same artifact: Pull the harness as a container for local-equivalent behavior docker pull mcr.microsoft.com/agents/harness:latest Or install the CLI directly dotnet tool install -g Microsoft.Agents.Harness A minimal agent definition wires the three things the runtime expects — chat client, instructions, tools — and lets the harness do the rest: js var agent = new HostedAgent { ChatClient = new AzureOpenAIClient env.OPENAI ENDPOINT, env.OPENAI KEY , Instructions = "You review PRs and flag risky diffs.", Tools = { new FileReadTool , new GitDiffTool } }; await agent.RunAsync "Review PR 482" ; For Foundry Hosted Agents, the deployment registers the same definition against the hosted runtime; consumption billing kicks in on first invocation. The local binary and the hosted binary share the same configuration surface — develop locally, promote to hosted without rewriting. For the GitHub Copilot and Claude SDK connectors, the connector packages register against the same governance policy endpoint as the harness itself. Approval rules, audit logs, and OpenTelemetry traces fan out across every model path the harness delegates to. Tool churn is the constant. Today's harness is Microsoft's; tomorrow's might be Anthropic's or someone's open-source rebuild. The model underneath the harness will keep getting cheaper, faster, smaller — that's the half of the AI story that's actually about AI. The other half — the 98.4% — is the part that doesn't move when the model does. Approval gates, history persistence, compaction, telemetry, policy enforcement, the loop that decides what gets called next. That's the layer worth investing in, and that's the layer OTF sits underneath: a single component surface where the same primitives look and behave the same on web, iOS, and Android, regardless of which model or harness is doing the thinking. Models change. Harness shapes change. The constraint that holds the system together — one API, one render path, one set of invariants across every surface your users touch — that's the durable part. Use the new Microsoft runtime. It's a real production story. And build the part that doesn't change when the runtime does. Microsoft's Agent Framework production runtime is the first time a major platform vendor has shipped the harness layer as a supported, GA product rather than a reference implementation. The 98.4% number reframes the whole agent conversation: the model is a swap-out part, the harness is the product. Foundry Hosted Agents handles deployment. The connectors handle model variety. The governance plane handles the audit story. And the part that survives all of it — the durable interface between your application and whatever agent runtime is hot this quarter — is the layer worth your engineering time.