Google’s Agent Substrate donation and the work that resumes after approval

Public developments and original analysis · Announcement date: 2026-11-05 · Sources

A quote is drafted on Friday afternoon, but its supervisor cannot approve it until Monday. If an agent keeps its execution resources throughout the wait, those resources are unavailable to other tasks. If the session stops, it may have to rebuild files and context on Monday. On November 5, Google announced its donation of Agent Substrate and gVisor to CNCF and described this kind of runtime problem. The gVisor project had already published its donation timeline in October. CNCF Sandbox is an early project stage. The execution sandbox is the mechanism that isolates code.

Agent Substrate separates an agent’s session from the workers that provide compute. An Actor can retain memory and disk state. When the turn finishes and the session goes idle, the runtime can save a snapshot and free the slot in a shared worker pool. A later request can restore the session onto an available worker. gVisor or a supported microVM provides isolation. For workloads with short bursts and long waits, the design changes what a platform team measures: the number of sessions actually executing at once, along with the storage and spare capacity required to bring them back.

Draft state and approval state

Take the quote as an illustrative scenario. The agent may already have gathered requirements, selected a price list and created a draft that still needs approval. A recoverable environment can preserve that work. A supervisor’s decision, by contrast, needs to identify a specific accessible draft version and record who approved it and when. This lets another person or process find the approved object even if the original runtime has been replaced.

Over the weekend, the customer might send a different quantity or the company might switch to a new price list. Restored memory can show what the agent saw on Friday. The application still needs to determine whether the new quantity changes price, lead time or the scope of approval. When the change requires a fresh quote, retain the original approval with its old version and send the new draft for review. If the price remains valid on Monday and neither the customer details nor the document have changed, the approved material can continue through the process.

These records serve different needs. The runtime can manage documents in progress and temporary tool state. The approved version and the result of customer communications should remain accessible in business records. If an external request times out after resumption, first query whether the quote was already sent or saved, then decide whether to retry. Otherwise a restart can result in the customer receiving two quotes. The application needs business records to establish the result, even when its code runs inside a sandbox.

Mokyun’s integration description relates the customer, inventory and price evidence for a quote to existing enterprise systems and calls for verifying versions, authorization and the recorded result. In a comparable design, a digital employee can prepare a quote package with data sources and hand it over through the enterprise’s existing approval and recording process. Choosing Agent Substrate as the runtime would depend on local infrastructure, storage and recovery tests. The published Mokyun description supports the division of work above.

When is keeping the session worthwhile?

If a task only needs a draft and a few fields while it waits, a database record can hold the pending work and a new execution can start after approval. Snapshots become more attractive when a session is expensive to reconstruct, for example after loading substantial dependencies or accumulating intermediate work, or when many similar sessions need to resume in quick succession. Snapshot size, version compatibility, retention and failed recovery all matter alongside wake-up time.

The distribution of waits matters as well. If most approvals arrive quickly, writing and reading a snapshot may be more work than keeping a process briefly. If approvals cluster on Monday mornings, the shared pool still needs enough capacity to handle that peak. Estimate simultaneous restores from actual approval times before comparing storage costs and idle capacity. An all-day average can hide the periods that drive provisioning.

A trial can compare staged execution with snapshot recovery on a set of past tasks with comparable waits. Measure time from approval to an actionable draft, work that must be repeated, and runtime and storage resources per completed task. Include a case where the source data changed and see whether the task proceeds with an old price. Platform teams can measure how quickly the session returns. Business teams can verify whether the next action still has a valid basis. Both results inform which tasks are worth keeping as complete sessions.

• Mokyun system integration

Mokyun Editorial · neo@mokyun.com