What context enters the run?
Local inputs and mounted resources shape what the agent can reach.
Read the research track →Agents carry context through workspaces, sessions, cached state, and tools. Across six Google VRP submissions, Kronowave investigated which authority reaches the next action when that context changes hands.
Six submissions across Google agent and developer surfaces share a single research thread: identify the context an action inherits, then compare it with the authority the caller actually supplied.
Local inputs and mounted resources shape what the agent can reach.
Read the research track →Hosted sessions keep working after one request has ended.
Read the research track →Cached objects can outlive the context in which they were created.
Read the research track →Tool routing and connections determine the final effect.
Read the research track →An agent may begin with one caller and continue through another execution component. The important record is the authority that arrived at the final effect.
Local inputs and mounted resources shape what the agent can reach.
Hosted sessions keep working after one request has ended.
Cached objects can outlive the context in which they were created.
Tool routing and connections determine the final effect.
The submissions use pinned implementations, healthy reference runs, changed-condition comparisons, and effect receipts to answer one authority question at a time.
Identify the implementation and runtime under examination.
Establish the expected authority and healthy effect.
Observe a specific handoff, reuse, or routing boundary.
Tie the effect to the actual caller and execution context.
Agent input may include a local workspace, mounted material, and an execution environment.
The research distinguishes the caller’s intended scope from the context made available to the running agent.
The workspace and resources associated with the authorized run.
The context the runtime used when processing the action.
A useful boundary record names both the initiating authority and the resources the run could exercise.
Hosted execution can retain work and state across individual requests.
The research follows session identity through continuation and checks which earlier authority the later action inherited.
The initiating request and its authority.
The actor and state associated with a later effect.
A long-lived session needs an account of who is authorizing each consequential continuation.
A cache or retained object may be consulted after the surrounding context changes.
The research compares the object’s origin with the authority active when it is used again.
The context that produced or stored the object.
The context and effect that followed its later use.
The owner of retained state is part of the authority of the action it helps produce.
Agent plans can pass through adapters, connections, and provider tools.
The research ties the selected route to the identity and effect recorded at the receiving boundary.
The authority and destination the caller intended.
The identity and result at the tool or provider.
Trace the authority to the system that performed the action.
An agent system has to preserve who can act as state moves between components. These Google submissions examine that continuity at four different kinds of boundary.
Six submissions to Google’s VRP anchor this family with source-pinned experiments, healthy controls, changed-condition comparisons, and execution receipts. The public synthesis shows the authority questions while the submitted reports carry the reproductions.
Tell us about the workflow you are shipping. We will start with the action you need to understand.
48-hour Rapid Scan →