Google updated its Gemini managed-agent tools with a new harness that brings its Antigravity coding agent's behaviour into its developer studio, alongside new interfaces for handling files and credentials. A "harness" here is the surrounding software that lets an AI model safely use tools, read files and remember context across steps.
The update reflects an industry realisation that the model is only half the product; the scaffolding that lets it act reliably is the other half.
What it means: when comparing AI tools, the underlying model matters less than how well the surrounding system manages files, permissions and memory. A modest model in a well-built harness often beats a stronger model bolted on clumsily, so judge the whole product, not the model name.