Parallel News · Edition No. 120 · Sat 24 Oct 2026

Every Agent Gets a Box.

Google, NVIDIA and VAST shipped isolated compute for each agent this week. That makes the sandbox a commodity, and the value moves to the gateway that decides what each agent may do and the record of why it did it.


26 monitors checked · 25 fresh ≤48h · 76 distinct signals · ~5 duplicates collapsed · 0 skipped

The Take — The sandbox is free now. Authorisation is the product.

In one 48-hour window, three different layers of the stack announced the same idea. Google Cloud unveiled Agent Compute Engine: every AI agent gets its own hardware-isolated microVM with a browser, a filesystem and network access, and Google says it already runs Gemini's agent work. VAST Data and NVIDIA shipped an Agent Sandbox Foundation Stack that keeps an agent's durable state apart from its throwaway compute, so a session can pause, resume or move between environments. NVIDIA added an OpenShell reference architecture, with BlueField-4 hardware enforcing policy out of band. Three of our monitors flagged Google's launch independently, the most corroboration any item got today.

When the biggest cloud, the leading GPU vendor and a storage company all ship the same primitive in the same week, the primitive stops being a moat. A per-agent box is becoming as ordinary as a container. The question that matters is what sits around the box.

The rest of the sweep answers that. BeyondTrust made AI Agent Security generally available on endpoints, separating agent activity from human activity and checking each action at runtime. Okta extended MCP governance across multiple servers, Datawiza launched an identity-aware Agent Gateway, and F5 is pairing its guardrails with Google's Agent Gateway. Palo Alto Networks set out five security workstreams with NVIDIA. The identity monitor logged eight items in a day: Ping, Saviynt and WorkOS all describe the same pattern. Carry every human-to-agent and agent-to-sub-agent hop inside the token, and authorise each tool call against both the user's permissions and the agent's.

Isolation limits how far a mistake can spread. Authorisation decides whether the action should happen at all. Only the second has pricing power.

Money arrived on the same rails. Stripe now lets agents pay invoices, move funds and convert currencies through Treasury. It uses agent-tagged API keys and approval rules, and live refunds and payouts can require a human to confirm. AEON extended agentic checkout to hotel bookings. Elliptic argued that spending limits and allowed counterparties have to be set before an agent transacts. Once an agent can move cash, the gateway's decision is a financial control.

The evidence on governance gaps is getting sharper too. Redwood's survey of 300 senior leaders found that 77% had already suffered a negative consequence from an ungoverned agent. Sopra Steria Next reports that fewer than 25% of production AI decisions can be traced and replayed end to end. Gen Threat Labs got an email-reading agent to run system commands and send data outside its environment. F5 catalogued coding agents leaking more than 13,000 internal screenshots.

The desk's proposal is a missing piece we will call a blast-radius warrant. It is a single-use, signed object issued by the gateway for one action. It binds the delegation chain, a reversibility class (read, reversible write, irreversible write, money), a spending ceiling and a time limit. The sandbox would refuse any system call that the warrant does not cover, and the spent warrant would become the audit record. Today the identity token, the sandbox policy and the decision trace live in three products from three vendors. Fusing them into one artifact is the part nobody has shipped.

The counterpoint deserves a hearing. Fine-grained, per-action authorisation could choke agents with latency and approval fatigue, and the CIO piece on "autonomy creep" shows how teams quietly remove human approvals once those approvals get in the way. The hyperscalers may also bundle gateway and audit into the sandbox, so independent vendors never get the chance. Google already pitches Agent Compute Engine together with "broader agent governance." If that bundle works, the value goes to whoever owns the compute after all.

The tie-breaker is cross-cloud. Agents increasingly call tools in other clouds, MCP servers, payment rails and other firms' agents. A warrant that only works inside one cloud does not cover that traffic. A portable one would.

Signals that matter

  1. Google Cloud Agent Compute Engine (Oct 23): a hardware-isolated microVM per agent. The sandbox is now a cloud product.
  2. Stripe Treasury for AI agents (Oct 22): agents can pay invoices, move funds and convert currencies, with agent-tagged keys and approval rules.
  3. BeyondTrust AI Agent Security GA; Datawiza Agent Gateway; Okta MCP governance (Oct 22–23): runtime, per-action authorisation of agents.
  4. Delegation chains in tokens (Oct 24): Ping, Saviynt and WorkOS describe the same pattern, which is close to a standard.
  5. VAST × NVIDIA Agent Sandbox Foundation Stack (Oct 22): durable agent state kept apart from throwaway compute.
  6. Redwood / Sopra Steria data (Oct 23): 77% hurt by ungoverned agents; fewer than 25% of AI decisions can be replayed.
  7. Gainsight drops per-seat pricing (Oct 23): prices on ARR plus accounts managed instead.
  8. Funding: Flapping Airplanes $1B at a $6B valuation; Blacksmith $66M Series C; Reduck $3M pre-seed for browser-automation MCP.

Regulation monitor stale since Oct 19 (Philippines HB 11014, AI Agent Transactions and Accountability Act). Significance rankings are editorial estimates, not investment advice.

← All Parallel News editions