Microsoft Agent Framework just collapsed multi-step agents from 28 seconds to 13. CodeAct changes the math.

Your agents are slow because every tool call is a separate model turn.

Microsoft just shipped CodeAct inside the Agent Framework.

Instead of: call tool → wait → call tool → wait → call tool

The model writes one Python script, calls your tools in a loop, and runs it once in a Hyperlight micro-VM.

Benchmark on multi-step workloads:

Traditional: 27.81s, 6,890 tokens
CodeAct: 13.23s, 2,489 tokens
52.4% faster. 63.9% fewer tokens.

Each tool call runs in an isolated micro-VM. One sandbox per invocation. Strong isolation is now free.

The package is agent-framework-hyperlight. Alpha. Available now.

One-line change to wire it:

codeact = HyperlightCodeActProvider(
tools=[get_weather, send_email],
approval_mode="never_require",
)

That is the entire integration surface.

This lands alongside two other pieces from BUILD 2026:

1. Agent Harness — production patterns built into the framework. Automatic context compaction so agents do not overflow mid-loop. FileMemoryProvider for session-scoped persistence. TodoProvider for multi-step task tracking. AgentModeProvider for plan vs execute separation. BackgroundAgentsProvider for parallel fan-out orchestration.

2. Foundry Hosted Agents — scale to zero, pay nothing while idle. Per-session VM isolation. Files and disk state persist across scale-to-zero. OpenTelemetry traces flow into Application Insights with zero extra wiring.

Agent Framework is 11.3k stars. 2,290 commits. AutoGen and Semantic Kernel are now one product.

If your agent orchestration is bottlenecked by tool-call latency, CodeAct is the architectural answer. Start there.

Audit your agent loop today. Count the model turns per task. That number is your latency floor.