# Microsoft Agent Framework Harness Is GA — The Runtime, Not the Library > Source: > Published: 2026-08-05 04:18:31+00:00 Microsoft’s Agent Harness and Foundry Hosted Agents hit general availability this week, completing a pivot the company has been telegraphing since Build 2026: the Agent Framework is no longer primarily a library for building agents. It’s a production runtime for governing them. That distinction matters. April’s 1.0 release gave you the SDK. This release gives platform teams the execution layer — tool calling, history, memory, context compaction, approval workflows, and OpenTelemetry observability — all on by default, all stable. GitHub Copilot SDK and Claude Agent SDK connectors are GA alongside it, meaning third-party coding agents now run as policy-governed fleet members, not one-off integrations. ## 98.4% Harness, 1.6% Model A VILA-Lab analysis of Claude Code from April 2026 found that 98.4% of the codebase consists of harness infrastructure — permissions, context management, sandboxing, tool routing, recovery logic — with only 1.6% dedicated to AI decision logic. That’s not a criticism of Claude Code. It’s the actual distribution of effort in any production agent system. The Agent Harness packages that 98.4%. According to principal software engineer Wes Steyn, “A model on its own can only generate text.” The harness is what turns text generation into autonomous task execution. Developers provide a chat client, instructions, and tools. The harness handles the rest through a single API call and runs as one binary across local development, containers, and Foundry Hosted Agents. ## Third-Party Agents Now Have a Governance Layer The more interesting story is the third-party connector architecture. The GitHub Copilot SDK and Claude Agent SDK connectors don’t just let those agents interoperate with Microsoft’s framework — they bring them under the same governance model. No custom adapters. Both inherit the Harness’s identity, content safety, and observability policies automatically, with traces routing to unified Foundry dashboards. For enterprise teams, this reframes the question. You’re no longer asking “can my agent run?” You’re asking “under which policy does it run, and where does the trace land?” That’s a fundamentally different operational posture — and a much more defensible one for teams accountable to compliance or security review. ## CodeAct: The Performance Story Worth Watching CodeAct is the other notable piece, though it comes with a caveat. Instead of sequential model-tool-model round trips, CodeAct lets the model write a single Python program that calls all necessary tools inside a Hyperlight micro-VM. On a representative multi-step workload (computing order totals across eight users), the results were meaningful: | Metric | Traditional | CodeAct | Improvement | |---|---|---|---| | Latency | 27.81s | 13.23s | 52.4% faster | | Tokens | 6,890 | 2,489 | 63.9% reduction | The caveat: CodeAct is still alpha. Python only, no macOS support yet, and approvals gate the entire `execute_code` block rather than individual tool invocations. For read-only, multi-step data workflows it looks genuinely promising. For anything involving side effects — payments, production writes, email — it’s not the right tool yet. [Microsoft’s documentation](https://devblogs.microsoft.com/agent-framework/codeact-with-hyperlight/) is explicit about this distinction. ## Multi-Agent Systems Without Rigid Pipelines The Handoff feature replaces fixed multi-agent topologies with dynamic routing. Instead of a predetermined “Planner → Implementer → Reviewer” chain, agents escalate, delegate, or complete based on actual conversation state. Combined with background sub-agents for parallel task execution, it’s the pattern that makes complex agent workflows maintainable without becoming brittle. ## Microsoft vs. Everyone Else The obvious comparison is LangGraph, which has spent 2026 focusing on durable execution, human-in-the-loop interrupts, and provider-agnostic portability. The choice is actually clear: Azure, .NET or Python, enterprise governance out of the box — Microsoft Agent Framework. Flexibility, provider independence, deep Python/TypeScript ecosystem — LangGraph. AutoGen entered maintenance mode in October 2025. The migration path is now well-defined. That’s the practical action item from this release for a meaningful slice of the .NET developer base. For .NET and Python developers building on Azure: the Harness eliminates substantial infrastructure work through built-in governance and observability. Evaluate CodeAct for read-only multi-step workloads when it exits alpha. Check whether your third-party agent setup benefits from running as governed fleet members. The full GA details are in the [InfoQ coverage](https://www.infoq.com/news/2026/08/agent-framework-harness-ga/) and the [Microsoft Agent Framework blog](https://devblogs.microsoft.com/agent-framework/microsoft-agent-framework-at-build-2026-announce/).