The Missing Link in Agent Infrastructure
Following its introduction of the Credentials API, which established a secure egress proxy for sensitive secrets, Google has released the antigravity-preview-09-2026 managed agent harness. This update introduces the Files API—a development that marks a structural shift in how agent environments handle data. By enabling persistent file and directory management, the Files API completes the sandbox I/O model, effectively formalizing the data-movement layer of the agent stack.
Completing the Sandbox I/O Model
Previously, agent environments were largely transient, often requiring full file rewrites that were both inefficient and token-heavy. The Files API changes this by allowing agents to upload, list, and download files directly into a persistent Linux sandbox. When an agent initiates its first interaction, the system creates a sandbox and returns an environment_id. Subsequent calls reuse this environment, transforming the harness from a simple execution node into a stateful data-processing environment.
This transition to stateful agents is measurable. According to Google’s documentation, the changes result in approximately 40% fewer output tokens on file edits and an 8% increase in task completion rates for multi-turn software engineering and research workflows. By moving from full-file rewrites to line-range replacements, the infrastructure now supports a more efficient lifecycle: ingest via the Files API, process via the Gemini 3.8 Flash model, and secure egress via the Credentials API.
A Layered Infrastructure Arc
The release of the Files API aligns with a broader industry trend toward standardizing the agent stack. As the ecosystem matures, distinct layers are emerging: AWS provides the compute runtime through AgentCore, Fastly manages edge governance, and Google is defining the standards for data movement and credential proxying. Meanwhile, the CNCF is evaluating the Model Context Protocol (MCP) as a potential cloud-native wire specification, and Apple’s Safari MCP continues to expand the protocol’s reach into browser-side contexts.
This convergence suggests that agent infrastructure is moving away from monolithic, execution-only environments toward interoperable, data-aware architectures. The ability to maintain state across tasks, combined with secure, non-resident credential handling, provides the foundation for agents to act as persistent intermediaries in complex, multi-turn workflows.
Cost and Scheduling Infrastructure
The Files API also arrives alongside new cost and scheduling primitives that complete the operational picture. Google’s release notes indicate that antigravity interactions follow a pay-as-you-go model, with estimated costs ranging from $0.30 to $1.00 for research and synthesis tasks, $0.70 to $3.25 for data processing, and up to approximately $5 for complex agentic workflows. Environment compute—CPU, memory, sandbox execution—is not billed during the preview period.
The new Triggers feature enables cron-based scheduling for agent tasks, allowing developers to bind an agent, environment, prompt, and schedule into a persistent resource that fires without manual intervention. Each execution reuses the same environment, so files created in one run persist and are visible to the next—reinforcing the stateful model the Files API enables.
Developer Migration and Breaking Changes
For developers currently utilizing the antigravity-preview-05-2026 harness, the transition to the 09-2026 version is mandatory. The older environment is scheduled for shutdown on October 5, 2026, creating a clear deadline for migration. This update introduces several breaking changes that require immediate attention.
The new API mandates the use of PascalCase for all parameters, replacing the previous snake_case convention. Additionally, developers must update their implementations to use the new file tool suite, which includes write_to_file, replace_file_content, view_file, find_by_name, grep_search, and list_dir. These tools are specifically optimized for the new line-range replacement model, which is central to the observed performance gains.
Open Questions for the Agent Economy
While the formalization of a data-movement layer is a clear step forward, several questions remain. It is currently unclear how persistent sandbox models will interact with evolving cloud-native security standards as they become more widely adopted. Furthermore, the industry has yet to determine whether the data-movement layer will eventually become a commodity service or remain tightly coupled to specific model providers.
As vendors continue to ship diverse credential architectures—ranging from Google’s egress proxy to Rubrik’s short-lived scoped tokens and Microsoft’s MCP Firewall—the challenge for developers will be navigating these competing standards. The focus for the coming months will likely remain on how these disparate layers—compute, governance, and data—can be unified into a cohesive, interoperable stack.