Google DeepMind Unveils Project Astra 2 -- Persistent AI Agent with Cross-Session Memory and Multimodal Awareness

## From One-Shot to Persistent The original Project Astra prototype showed real-time multimodal awareness -- identifying objects through a camera and answering questions about the physical environment. Project Astra 2 extends that foundation into genuinely persistent operation: the agent remembers prior conversations, tracks ongoing tasks, and resumes them after interruption. DeepMind's demonstration showed an agent assigned to monitor a procurement pipeline over three days. It checked supplier portals each morning, flagged price anomalies, composed draft messages to vendors, and summarised status each evening -- without a single human restart command. ## Sub-Agent Coordination Astra 2 introduces a layered architecture in which the primary agent spawns and coordinates specialised sub-agents: a web-browsing agent for data retrieval, a document-generation agent for drafting, and a code-execution agent for data transformation. Results flow back to the orchestrator, which maintains a task graph and decides next steps. This architecture closely mirrors how human teams divide analytical and execution work -- and signals that the practical unit of AI deployment is shifting from a single model call to a multi-agent workflow. ## Commercial Timeline Google has not announced a public API release date for Astra 2 capabilities. A limited preview is available to select Cloud customers through Google AI Studio. Full integration with Google Workspace is expected in Q4 2026.