Google Cloud maps when enterprise agents earn autonomy

Anna Whelan and Amulya Saridey laid out how enterprises move AI from answering to acting, in Google Cloud Events session on architecting an agentic center of enablement.

The punchline came early: maturity is not a product you buy.

Whelan, an AI consultant at Google Cloud Consulting, and colleague Amulya Saridey said they built the framework after work with large enterprise customers moving from experiments to production. They frame maturity as trust and delegation, the same way you test a new teammate on low risk tasks first, and they measure it across three levers that scale together: reasoning depth from single turn answers to chain of thought planning and self correction, architectural complexity from passive RAG memory to tool agency and multi agent networks, and decision making from reactive checklists to proactive intent with human in the loop safeguards.

Level one is the sidekick. It is reactive, stateless, built on retrieval augmented generation and generalist integrations, good at summarizing a fifty page vendor contract or searching wikis but it forgets the why once the tab closes. The team said most enterprises live here today. It makes individuals faster without changing the process.

Level two turns the assistant into a workflow agent. The trigger shifts from a prompt to a business event like a new ticket, a price threshold, or an IoT stock error, and the agent uses chain of thought to decompose a goal, call tools via APIs, and keep task specific persistence until the job is done. Humans stay in the loop to approve the final action, but the KPI changes from hours saved to cycle time reduced.

Level three is collaborative autonomy. The architecture becomes multi agent, often a supervisor pattern or peer to peer network, with a defining A2A handshake where agents negotiate, share context, and hand off subgoals without human mediation. Long term memory carries preferences across sessions and nonlinear orchestration lets agents find a new path and self correct when blocked. Humans move from in the loop to on the loop, supervising outcomes not clicks.

Level four is the adaptive system the team describes as proactive and predictive. It runs on a semantic data fabric that gives agents unified access to truth, plus closed loop reflection that lets agents critique and optimize their own logic in real time. Governance shifts to safety reflexes, hard coded guardrails that bypass the model to kill an action that would violate security, ethics, or compliance.

The roadmap gets concrete in the transitions. To go from one to two, the consultants prescribe four moves: pick high friction, repeatable processes that can be mapped to discrete steps, implement tool calling into CRMs and ERPs so the agent gains transactional hands, establish persistence for a job lifecycle, and define success as measurable operational metrics.

The jump from two to three is the biggest structural leap. It requires breaking a monolith into specialized agents, standardizing on open protocols like A2A to avoid a tangle of custom code, adding long term memory for cumulative intelligence, and formally shifting governance to delegated autonomy.

To reach four, they add reflection loops that learn from live runs, a unified semantic layer that dissolves silos, automated governance via those safety reflexes, and proactive sensing so agents initiate fixes on environmental signals before a human files a request.

The fictitious retailer Omni Retail Global, twelve hundred stores plus global e-commerce, makes the abstract tangible. At level one it used RAG to summarize contracts. At level two it gave an inventory agent API access to SAP ERP and FedEx tracking, with persistence on a job ID that followed an anomaly for three days and a chain of thought loop that verified sensor data, checked sales logs, and if the gap exceeded five units, called the warehouse system to trigger a cycle count for bin 402.

At level three, a viral skincare spike shows the collaboration. A social listening agent signals procurement, a supervisor delegates to agents that secure factory stock, reroute trucks, and update e-commerce pricing, and when freight fails the agents renegotiate the hold with the supplier and find a backup route. The result they cite is shelves restocked in forty eight hours while competitors were still clearing silos. For level four, the same surge feeds a reflection loop that learns an alternate port would have saved two thousand dollars in freight and rewrites its routing logic, while a semantic fabric links weather, celebrity posts, and fuel prices and safety reflexes block any transaction over one million dollars or a labor compliance breach.

That last part is where outside context helps. A separate enterprise forecast from AMCAP Global this week describes 2027 agentic maturity as hinging on Model Context Protocol standards, Model Agnostic Agent Orchestration, and Confidential Virtual Machine execution environments reaching enterprise grade. That framing centers protocol and confidential compute, while the Google Cloud Events framework centers A2A negotiation and hard coded safety reflexes as the governance spine. The contrast matters for buyers choosing where to standardize first.

Not every workflow needs to climb to four. The consultants repeat that point at each stage, and it is the most useful constraint in the deck. If your process cannot be mapped to clear steps, instrumented with APIs, and measured on a hard KPI, you do not have an agent yet, you have a faster search box.