Google has introduced Agent Executor, an open source runtime aimed at helping enterprises run AI agents more reliably at scale, as attention shifts from building agent prototypes to managing the operational challenges of putting them into production.

To address those production-related challenges, the runtime, according to the company, comes with capabilities that are geared towards supporting long-running and distributed agent workflows.

Typically, long-running agent workflows are AI-driven tasks that execute over extended periods, from minutes to days, often involving multiple steps, system interactions, pauses for human input, or recovery from interruptions before reaching completion.

For such workloads, the runtime includes support for durable execution, allowing workflows to resume after outages or human approvals, along with secure sandboxing for isolating agent components, session consistency controls for distributed workflows, and connection recovery features intended to preserve execution state during network interruptions, Google wrote in a blog post.

The runtime also supports “trajectory branching,” which allows developers to test alternate execution paths from saved checkpoints without losing prior context, it added.

Furthermore, Agent Executor bridges multiple deployment models, including on prem and pre-built or custom managed agents, the company said, allowing users to mix and match between any or all of Google Antigravity, frontier agents built by Google, agents built by the user and managed by Google, and custom agents and agents using Agent2Agent (A2A) protocol, as desired.

Targeting production reliability gaps

Analysts and experts see value for both developers and enterprises in Agent Executor’s capabilities.

“Durability, orchestration, and resumability are the real blockers for any enterprise production agents,” said Advait Patel, senior reliability engineer (SRE) at Broadcom.

“What kills enterprise adoption is agents that lose their state when a pod restarts, sessions that corrupt under concurrent writes, or long running workflows that cannot recover from a network blip. Once your agent is taking actions on real systems, you cannot afford it to forget what it did halfway through,” he pointed out.

“The event log, snapshotting, single writer model, and connection recovery in Agent Executor are exactly the things SRE teams have been duct taping for the last year,” Patel noted, adding that existing frameworks such as LangChain and AutoGen are great for prototyping, but more often than not fall apart in production once agents run for hours or days.

For CIOs, said Gaurav Dewan, research director at Avasant, the runtime’s operational safeguards such as secure sandboxing, and checkpointing could prove just as significant for incident analysis and auditability.

At the same time, he cautioned that the runtime’s capabilities alone do not solve the broader governance and oversight challenges that CIOs continue to face with enterprise AI deployments.

 “Issues such as accountability, explainability of agent decisions, policy enforcement, and secure access across interconnected systems are still evolving,” he said. “As a result, while distributed runtimes can strengthen the operational backbone of agent deployments, CIO-level considerations around trust, compliance, and enterprise control are likely to require additional governance and oversight layers beyond runtime infrastructure alone.”

Using infrastructure layer to gain strategic advantage

Google, however, is not alone in trying to shape the emerging infrastructure layer for enterprise AI agents. Other hyperscalers, such as Microsoft, with AutoGen and AWS, with Bedrock AgentCore, are promoting open or interoperable frameworks to gain strategic advantage.

“There are growing indications that hyperscalers are converging toward a model that combines open or interoperable tooling at the top of the stack with monetization concentrated in underlying infrastructure layers,” Dewan said.

“Google, Microsoft, and AWS are increasingly offering SDKs, agent frameworks, and orchestration tools to drive developer adoption and ecosystem growth, while continuing to generate value through compute infrastructure, managed AI platforms, data services, and observability capabilities,” he added.

And, according to Patel, Google’s strategy around Agent Executor is reminiscent of the path that the hyperscaler followed with Kubernetes ten years ago: “Give away the runtime, [and] drive consumption on Google Cloud via services, such as the Gemini Enterprise Agent Platform and Managed Agents API.”

He added, “[hyperscalers] have figured out that proprietary agent frameworks will not get adopted at enterprise scale. The money is in cloud consumption, managed services, and model inference. The tools on top need to be open or nobody will trust them.”