Agent computer / working thesis
View context.md

Stateful serverless workers

Persistent identity. Ephemeral computer.

An agent should not be a permanently running VM. It should be a durable organisational worker that leases a computer only while work is happening.

Working architecture noteResearch snapshot: 17 August 2026Capacity figures are illustrative
Agent seatIdentity · role · policy
WorldMemory · state · lineage
Computer leaseTemporary execution

The conceptual shift

The model is not the employee.

A chatbot waits and answers. A copilot assists a human. RPA follows a predefined flow using borrowed credentials. A digital worker owns bounded outcomes, preserves continuity and escalates exceptions.

The employee-shaped thing is not a model endpoint. It is the complete governed deployment:

Digital worker = model + harness + memory + World
               + computer + organisation identity
               + bounded authority + role + supervision + audit

The model supplies cognition. The harness supplies working method. Memory supplies experience. The computer supplies a desk and hands. Organisational identity supplies membership. Capabilities supply keys. Policy defines authority. Human takeover supplies managerial intervention.

The employee is not the computer. The employee borrows a computer when work arrives.

Stateful serverless workers

Asleep does not mean absent.

A worker can retain its identity, memory, inbox, files, pending work and schedule while consuming no CPU or RAM.

SLEEPING → event → WAKING → computer leased → WORKING → checkpoint → SLEEPING

Microsoft Graph notifications, Gmail push events, Teams mentions, SharePoint changes, calendar events, CRM webhooks and schedules wake the worker. The control plane receives events. An idle desktop does not need to poll Outlook all night.

Working definition

A stateful serverless worker is a durable organisational actor whose execution environment exists only while it is doing work.

Waiting for a person should also release compute

If an agent stages a £40,000 quotation and asks for approval, it should not hold a VM for six hours while a director is in a meeting. It checkpoints the World, releases the computer and wakes when Teams delivers the decision.

Action staged → World checkpointed → computer released
        ↓
Approval in Teams
        ↓ three hours later
Manager approves → worker wakes → work continues

Concurrency, not headcount

Ten thousand workers do not need ten thousand VMs.

The capacity question is not how many workers exist. It is how many are active at the same time.

average active = provisioned workers × active minutes/day ÷ 1,440

Model the workforce

Illustrative concurrency arithmetic only. Real density depends on workload, OS, burst shape, browser use, model latency and API quotas.

69Average active computers
209Peak computers to provision
14–2764 GB browser hosts at 8–16 active workers each

10,000 × 10 ÷ 1,440 = 69.4 average concurrent

Dormant means zero CPU and RAM, not zero cost. Storage, queues, identities, subscriptions, audit and scheduling remain. API-only work can reach extreme density; full Chromium desktops are moderate; Windows Office and GPU workloads are lower.

Cold

Durable World only. Zero CPU/RAM. Wake in seconds.

Warm

Snapshot cached on fast storage. Near-instant resume target.

Live

Computer allocated. Display and takeover available.

Durable objects

Separate the worker from its body.

The worker feels persistent to the company, but models and computers remain replaceable.

SeatRole, organisational identity, permissions, supervisors and continuity.
WorldFiles, browser/profile references, task state, history, checkpoints and lineage.
ComputerReplaceable Linux or Windows execution capacity acquired through a lease.
Org packM365, Google Workspace or another provider's identity and application adapters.
BrokerHost-side scoped authority. Reusable organisation credentials never enter the guest.

Generic runtime pools serve browser, Office, engineering and media workloads. A task attaches one encrypted World overlay to one clean computer, grants short-lived capabilities, checkpoints the result and destroys or resets the runtime.

Concurrency is explicit. A worker processes its mailbox serially by default; parallel tasks receive independent copy-on-write branches rather than racing over one browser profile.

Provider neutrality

The computer should know almost nothing about Microsoft or Google.

Organisation-specific identity and application access belong in pluggable Organisation Packs and a capability broker outside the guest.

Microsoft 365

Graph for mail, calendar, SharePoint, OneDrive, Teams and directory work. Entra OBO for delegated use; approved agent identity or managed identity for autonomous work. Windows 365 where desktop Office or compliant-device workflows require it.

Google Workspace

Gmail, Drive, Docs, Sheets, Calendar and Chat APIs through delegated OAuth or workload identity. Domain-wide delegation only where explicitly approved. Browser automation remains a fallback.

Expose common operations where the overlap is real—mail reading, drafting, file search and calendar work—while preserving native provider capabilities. Agnosticism should not erase what makes either ecosystem useful.

Agent computer → provider-neutral capability request
               → tenant broker → short-lived token
               → Microsoft Graph or Google Workspace

Human co-presence

Teams becomes the supervisor cockpit.

Takeover is not a remote-desktop gimmick. It is an explicit, exclusive and audited control lease.

  1. The worker pauses and sends a private Adaptive Card: Watch, Take control, Deny or Stop.
  2. The action opens a Teams personal tab or dialog.
  3. Teams SSO identifies the tenant and operator.
  4. The control plane validates the supervisor and issues a one-use viewer/control token.
  5. An embedded WebRTC viewer displays the real computer.
  6. Taking control stops agent input. The human resolves MFA, CAPTCHA or judgement-heavy work.
  7. On release, the agent receives the resulting application state—never the password or secret input.
AGENT_CONTROL → HANDOFF_REQUESTED → HUMAN_CONTROL
              → HUMAN_RELEASED → AGENT_RESUME

View and control are separate permissions. Messages contain no bearer URLs or sensitive screenshots. Every pause, viewer connection, lease, release and resume enters the event history.

What exists today

The market has built the pieces separately.

Consumer agent computers, Microsoft Cloud PCs, hosted sandboxes, microVMs and orchestration projects each solve part of the system. None should be copied wholesale.

ProjectWhat it is doing wellWhat remains
Grok BotPersistent cloud computer, browser sessions, live observation and takeover. Best consumer interaction reference.Account-wide computer shares files, cookies and CLI credentials; separate screens are not security boundaries.
Windows 365 for AgentsPools, checkout/checkin, computer operation, WebRTC observation/takeover and Microsoft governance.Microsoft-native; sessions reset on release, so portable durable World continuity remains external.
DeepSeek HarnessAppend-only session truth, coherent capability seams, per-call policy and fail-closed enforcement.An agent harness rather than the isolated microVM substrate.
E2BEstablished hosted sandbox API, templates and broad agent-framework integrations.Hosted-first; substantial cloud infrastructure to self-host.
CubeSandboxProduction-shaped microVM fleet, egress controls, credential injection and snapshots.Heavy Linux/KVM infrastructure for an early single-host product.
Microsandbox / BoxLiteStrong local embeddable microVM candidates with lifecycle, persistence and snapshot direction.Need real Apple Silicon/Linux benchmarks and hostile-guest testing before choosing.
AIO Sandbox / SmolVMCoherent browser, desktop, shell and files experience.UX references rather than the final security boundary.
Kubernetes Agent SandboxTemplate → WarmPool → Claim model, stable identity and later fleet orchestration.Orchestrates but delegates isolation; premature before one-host lifecycle is solid.
NehemiahClean product abstraction: Sandbox = Machine; Computer = Machine + display.Early rather than a complete general-purpose substrate.

The opening

Composition is the product.

The underlying primitives are inspired by existing systems. The differentiation is the opinionated boundary around a provider-neutral SMB agent seat.

Grok interaction
+ Microsoft governance
+ Cube security
+ DeepSeek truth
+ Git state
+ Sym continuity
= provider-neutral agent seat

The moat is not creating VMs, streaming desktops, wrapping Graph, exposing MCP or snapshotting disks. It is cross-provider identity and access, persistent World semantics, human-agent supervision, governed external consequences and accumulated organisational memory.

Models provide intelligence. The agent seat turns intelligence into an accountable member of the organisation.

The best possible computer

Git for computers, not VNC with an API.

The ideal computer is a versioned, instrumented execution World that both agents and humans can inhabit.

  • One coherent filesystem, process, browser and desktop state.
  • API → shell/files → DOM/CDP → accessibility → pixels control hierarchy.
  • Verified read-only base + role layer + tenant configuration + encrypted World overlay + task branch.
  • Checkpoint, fork, diff and restore with explicit disk/live semantics.
  • Low-latency watch, annotation, exclusive takeover and emergency stop.
  • Default-deny egress, metadata/private-network blocks and bounded resources.
  • Host-side credential and capability brokering.
  • Append-only causal event history covering model-visible facts and external receipts.
  • No model or harness dependency in the machine core.
  • Truthful capability reporting with no silent degraded isolation.

The seat owns the durable World. Disposable Linux or Windows computers execute it. The same worker can change model, machine host or organisation provider without surrendering its continuity.

Build discipline

Non-negotiables.

  1. No raw long-lived organisation credentials in guests.
  2. No silent unconfined fallback.
  3. No default metadata or private-network access.
  4. No split local-file and remote-shell execution world.
  5. No unbounded process, output, disk, time or connection growth.
  6. No viewer without current machine and lease authorisation.
  7. No model or harness dependency in the machine core.
  8. CLI and SDK first; MCP optional.
  9. No fleet control plane before one-host lifecycle, cleanup and recovery are reliable.
  10. Every irreversible external effect must be policy-checked and receipted.

Commercial shape

Agent seat       identity, memory and governance
Computer usage   active execution
Storage          World and history
Premium runtime  Windows, GPU or high-memory work

The customer provisions a persistent governed seat and pays for execution actually used—not a continuously running VM per agent.

Primary references

Further reading.

The worker persists. The computer appears only while work is happening.

Ten thousand digital workers need durable state, an event mailbox and access to a computer when work arrives—not ten thousand permanently running machines.

Context copied. Paste it into any agent.