Microsoft's Agent Framework Harness Moves from SDK to Production Runtime

Microsoft has transitioned Agent Framework from an experimental SDK into a production-ready system with a governed runtime. Following the framework's 1.0 general availability in April 2026, the company shipped the Agent Harness and managed Foundry Hosted Agents at Build 2026 in June, now offering teams both the tools to build agents and the infrastructure to run them safely at scale.

What the harness does#

The harness provides the execution layer that sits between a language model and the outside world. A model alone generates text; the harness wraps it to enable tool invocation, multi-step reasoning, persistent state, and graceful task completion. Microsoft's release bundles function invocation, history persistence, context compaction, a planning-and-execution framework, file memory, skill composition, web search, tool approval workflows, and OpenTelemetry instrumentation—all enabled by default and individually configurable. More sensitive features like shell access and background sub-agents remain opt-in with warnings.

The framework runs as a single binary across development machines, containers, and hosted environments, meaning teams deploy the same code path from prototyping to production.

Why the harness matters more than it appears#

Research by MBZUAI's VILA-Lab quantified why infrastructure matters. An April 2026 analysis of Claude Code's briefly-exposed source estimated that roughly 98.4% of the codebase handles harness infrastructure—permissions, context management, sandboxing, tool routing, error recovery—versus 1.6% dedicated to AI reasoning logic. The finding carries caveats: it relies on a leaked bundle with generated and minified code, not a direct audit. Yet the direction aligns with independent convergence: agents built separately (Codex CLI, Aider, and others) arrived at the same harness architecture, suggesting it reflects problem constraints rather than arbitrary design.

A recent benchmark reinforced the point. Testing Agent Framework against GitHub's Copilot SDK under identical conditions—same model, deterministic mock environment—found both reached the same conclusions in the same reasoning steps. The differences lived purely in engineering: one harness self-enforced a 40-round-trip safety limit and exited cleanly; the other continued to 300 trips without built-in brakes, relying on the host to stop it.

Governance and observability at scale#

The coding-agent connectors exemplify Microsoft's focus on control. Agent Framework can orchestrate GitHub Copilot SDK and Claude Agent SDK agents without custom bridges; each retains its own loop while participating in a broader workflow. The critical detail is that traffic from third-party agents flows into the same identity, content safety, and observability policies as the rest of the fleet. Logs land in unified OpenTelemetry traces and Foundry dashboards rather than isolated integrations, shifting governance from "what can this agent do?" to "who launched it, under which policy, and where's the audit trail?"—a control pattern also emerging in reference architecture.