Microsoft's CodeAct Cuts Agent Latency 52% and Token Use 64% by Having the Model Write One Program Instead of Chaining Tool Calls

InfoQ·medium signalXBlueskyLinkedInCopy link

Microsoft Agent Framework's CodeAct has the model write a single short Python program that calls tools via call_tool(), executes it once in a fresh Hyperlight micro-VM per call, and returns a consolidated result — shipping in the alpha agent-framework-hyperlight package. On Microsoft's representative multi-step workload, traditional tool wiring took 27.81 seconds and 6,890 tokens while CodeAct took 13.23 seconds and 2,489 tokens: 52.4% faster and 63.9% fewer tokens. Reporting on the Agent Harness GA also cites an analysis that ~98.4% of the Claude Code codebase is harness infrastructure and only ~1.6% is AI decision logic.