# Desktop Agent／AI-RPA編

CURRENT_STATUS: UNKNOWN
OBSERVATION: SPEC_ONLY
FRESHNESS: UNKNOWN
PAGE_TYPE: reference
CANONICAL_URL: /tools/ea-050-desktop-agent-ai-rpa
CLAIM_IDS: none
EVIDENCE_IDS: EVD-EA-050-URL-800526331074, EVD-EA-050-URL-3CAB0D84B00C, EVD-EA-050-URL-4DDE3CD11E3A, EVD-EA-050-URL-3ED7A6CAE4A3, EVD-EA-050-URL-C5F3CD81FD74, EVD-EA-050-URL-580B8A0A8C2D, EVD-EA-050-URL-D7F007864A6D, EVD-EA-050-URL-51BC8AA1690C, EVD-EA-050-URL-FFBEF51B3785, EVD-EA-050-URL-DD3C5207FF0D, EA-050-R11-RESEARCH, EA-050-R11-WEB-READBACK
UNKNOWN_IDS: EA-050-UNKNOWN-01, EA-050-UNKNOWN-CURRENTNESS
OPS_IDS: none

## Summary
「Desktop Agent／AI-RPAは、従来RPAをAIで置き換えた単一カテゴリではない。API、structured UI automation、vision/reasoning computer useは異なる失敗様式を持つ実行経路であり、同じidentity・policy・machine isolation・audit境界の下で使い分ける。GUI操作が可能であることは、業務ルール上正しいこと、権限上許可されること、再現可能であることを意味しない。」

## Current judgment
- Desktop automation has at least two execution paradigms: (A) selector/UIA/action-graph based deterministic RPA and (B) screenshot/vision/reasoning based CUA. They can share host/session/credential infrastructure but should not share the same reliability assumptions.
- Windows UI Automation provides structured control semantics; CUA can act when selectors/APIs are unavailable or UI changes, but model reasoning introduces probabilistic action selection. Future Core should therefore prefer API/structured automation where available, deterministic UI automation next, and CUA as a bounded fallback or explicitly governed execution path rather than treating CUA as a drop-in replacement.
- unattended foreground automation is an identity-and-session problem, not merely a model capability problem. Microsoft and UiPath both require execution infrastructure and user/session credentials or equivalent non-human identity configuration for foreground UI interaction.
- Dedicated execution machines/VMs are operationally meaningful. Microsoft explicitly recommends dedicating a machine for computer use; same-machine jobs queue sequentially. Shared interactive desktops create interruption, state leakage, contention and attribution problems.
- OS privilege boundaries remain effective constraints. Windows UIA cannot simply cross protected/high-integrity UI by default. AI reasoning does not remove UAC/uiAccess boundaries.
- conventional RPA is not immune to platform drift: a 2026 Windows security update broke credential-dialog automation. Therefore deterministic selectors are more reproducible than CUA, but not invariant to OS/security changes.

## Scope
Enterprise AI Core / provider mapping / organization overlay / governance / benchmark / implementation / non-LLM component

## Confirmed
- Desktop automation has at least two execution paradigms: (A) selector/UIA/action-graph based deterministic RPA and (B) screenshot/vision/reasoning based CUA. They can share host/session/credential infrastructure but should not share the same reliability assumptions.
- Windows UI Automation provides structured control semantics; CUA can act when selectors/APIs are unavailable or UI changes, but model reasoning introduces probabilistic action selection. Future Core should therefore prefer API/structured automation where available, deterministic UI automation next, and CUA as a bounded fallback or explicitly governed execution path rather than treating CUA as a drop-in replacement.
- unattended foreground automation is an identity-and-session problem, not merely a model capability problem. Microsoft and UiPath both require execution infrastructure and user/session credentials or equivalent non-human identity configuration for foreground UI interaction.
- Dedicated execution machines/VMs are operationally meaningful. Microsoft explicitly recommends dedicating a machine for computer use; same-machine jobs queue sequentially. Shared interactive desktops create interruption, state leakage, contention and attribution problems.
- OS privilege boundaries remain effective constraints. Windows UIA cannot simply cross protected/high-integrity UI by default. AI reasoning does not remove UAC/uiAccess boundaries.
- conventional RPA is not immune to platform drift: a 2026 Windows security update broke credential-dialog automation. Therefore deterministic selectors are more reproducible than CUA, but not invariant to OS/security changes.

## UNKNOWN
- EA-050-UNKNOWN-01: provider横断でCUA action trace schemaを標準化できるか、business-rule violationを一般化して検出する評価標準、各providerのscreenshot/action retentionとtenant-controlled redaction差分、CUAとdeterministic RPAの自動fallback/rollback品質。
- EA-050-UNKNOWN-CURRENTNESS: 外部根拠の再確認日がcanonical body/control-planeで確定していないため、currentnessはUNKNOWNとして保持する。日付を推測補完しない。

## Evidence
- EVD-EA-050-URL-800526331074 (RESEARCH): 1) official product docs | "Automate web and desktop apps with computer use" | Microsoft | — https://learn.microsoft.com/en-us/microsoft-copilot-studio/computer-use
- EVD-EA-050-URL-3CAB0D84B00C (RESEARCH): 2) official product docs | "Configure where computer use runs" | Microsoft | — https://learn.microsoft.com/en-us/microsoft-copilot-studio/configure-where-computer-use-runs
- EVD-EA-050-URL-4DDE3CD11E3A (RESEARCH): 3) official product docs | "Run unattended desktop flows" | Microsoft | — https://learn.microsoft.com/en-us/power-automate/desktop-flows/run-unattended-desktop-flows
- EVD-EA-050-URL-3ED7A6CAE4A3 (RESEARCH): 4) official product docs | "Robots" | UiPath Orchestrator | — https://docs.uipath.com/orchestrator/automation-cloud/latest/user-guide/about-robots
- EVD-EA-050-URL-C5F3CD81FD74 (RESEARCH): 5) official product docs | "Configuring automation capabilities" | UiPath Orchestrator | — https://docs.uipath.com/orchestrator/automation-cloud/latest/user-guide/configuring-automation-capabilities
- EVD-EA-050-URL-580B8A0A8C2D (RESEARCH): 6) official platform spec/docs | "UI Automation Specification" | Microsoft Windows | — https://learn.microsoft.com/en-us/windows/win32/winauto/ui-automation-specification
- EVD-EA-050-URL-D7F007864A6D (RESEARCH): 7) official security docs | "UI Automation Security Overview" | Microsoft | — https://learn.microsoft.com/en-us/dotnet/framework/ui-automation/ui-automation-security-overview
- EVD-EA-050-URL-51BC8AA1690C (RESEARCH): 8) official failure report | "Fix: Power Automate desktop Windows credential dialog issues" | Microsoft | — https://learn.microsoft.com/en-us/troubleshoot/power-platform/power-automate/desktop-flows/windows-credentials-dialog-issues
- EVD-EA-050-URL-FFBEF51B3785 (RESEARCH): 9) primary research + author-maintained implementation | "OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments" | Xie et al. / xlang-ai | — https://github.com/xlang-ai/OSWorld
- EVD-EA-050-URL-DD3C5207FF0D (RESEARCH): 10) primary research implementation | "OS-Blind: The Blind Spot of Agent Safety" | lime-nlp | — https://github.com/limenlp/OS_Blind
- EA-050-R11-RESEARCH (RESEARCH): EA-050 監査済みresearch / canonical revision binding
- EA-050-R11-WEB-READBACK (IMPL): EA-050 Web body identity / canonical readback

## Failure / limits
Vendor statements that CUA "adapts" to UI changes are product claims, not guarantees of task correctness across arbitrary layouts.
## Failure / limits
OSWorld/OSWorld-Verified evaluates bounded benchmark tasks; scores do not establish production SLA, security, or success on organization-specific apps.
## Failure / limits
Desktop agents can act correctly at the GUI level yet violate hidden workflow/business rules; application semantics must be separately validated.
## Failure / limits
State hazards include modal dialogs, focus theft, resolution/DPI variance, localization, screen locking, concurrent user sessions, stale credentials, pop-ups, remote-desktop differences and application updates.
## Failure / limits
Security hazards include credential phishing, data exfiltration, destructive OS actions, malware/code execution, deceptive agreements and prompt/instruction content rendered inside applications. OS-Blind demonstrates these as evaluation categories; it does not prove prevalence in any particular product deployment.
## Failure / limits
A robot/service identity with correct infrastructure permissions can still hold excessive application-level permissions. Orchestrator/machine authorization and target-application authorization are separate gates.
## Failure / limits
EA-050初回Packetとして、Core構造、current provider実装例、OS privilege/session failure、benchmark/safety反証、availability/cost gates、後続再確認点が揃ったため停止。次回はEA-051へ進む。

---
## Practice
Microsoft Copilot Studio computer use: GA announced May 2026; requires generative orchestration; model choice and Copilot Credit billing are current/product-volatile. BYOM requires supported Power Automate for desktop version and machine registration/configuration.
## Practice
Microsoft unattended desktop flows: Windows edition/session/RDP/NLA/Entra configuration constraints apply; sovereign cloud/hosted capacity limitations vary by feature.
## Practice
UiPath: foreground unattended UI requires appropriate machine login credentials and configured robot/non-human identity; background Session 0 differs materially.
## Practice
Protected Windows UI: uiAccess/signing/install-location/UAC conditions can block automation even when ordinary application UI is accessible.
## Practice
MATURE for deterministic enterprise RPA; DEPLOYED/GA but rapidly evolving for CUA-assisted desktop automation. Recommended implementation pattern for future book: task router -> API/connector path when available -> deterministic RPA/UIA path -> governed CUA fallback; all paths terminate in common identity/policy/audit controls. Separate attended, unattended-foreground, background/session-0, and hosted/dedicated-machine profiles.
## Practice
Use OSWorld/OSWorld-Verified for cross-application capability evaluation and OS-Blind for safety evaluation. Do not merge capability score with safety score. For organization acceptance, create a private evaluation corpus from real applications with: deterministic completion checks, forbidden-action checks, business-rule assertions, recovery/retry behavior, resolution/localization variants, and clean-state reset between runs.
## Practice
least-privilege robot identity; dedicated machine/VM; prohibit unnecessary admin/uiAccess; policy-gate credential/system/security-setting changes; treat on-screen text as untrusted input for CUA.
## Practice
公開後本文では「desktop execution hierarchy」をCore化し、API > structured deterministic automation > governed CUAという優先経路を候補にする。
## Practice
provider mappingではMicrosoft Copilot Studio/Power AutomateとUiPathを、機能比較ではなくidentity/session/machine/queue/audit軸でマッピングする。
## Practice
publication時にMicrosoft computer-use GA、model lineup、credit rate、Power Automate minimum version、UiPath current Automation Cloud docsを再確認する。
## Practice
企業向けTEST_PROFILEを作る場合は専用VM、無害なsynthetic app/data、non-admin test identity、network allowlist、clean snapshot restoreを用い、real credentials/production tenantは使わない。
## Practice
Desktop Agent/AI-RPA is layered execution design; API, deterministic UI automation and vision/reasoning CUA have distinct failure modes. Model lists, credits, hosted-machine details and tenant availability remain PRODUCT_VOLATILE. GUI action capability does not establish business-rule correctness, authorization or reproducibility. Preserve all OPEN_UNKNOWN.
## WHY_THIS_MATTERS_TO_UPDATE
EnterpriseのGUI自動化は、従来RPAの決定論的selector/UI Automationと、Computer-Using Agent（CUA）の視覚・推論ベース操作が同じdesktop surfaceへ収束しつつある。両者は失敗様式、権限境界、credential/session要件、監査可能性、コスト構造が異なるため、公開後アップデートでは「AI-RPA」を単一機能として扱わず実行層を分離する必要がある。

## SCOPE_ROUTING
Enterprise AI Core / provider mapping / organization overlay / governance / benchmark / implementation / non-LLM component

## CLAIM_OR_RESEARCH_QUESTION
1. Desktop Agentと従来RPAを同一の運用モデルで扱えるか。
2. unattended GUI executionで最低限分離すべきidentity/session/machine/privilege境界は何か。
3. GUI agent benchmarkの成功率を業務自動化の信頼性へ外挿できるか。
4. deterministic UI automationとvision/reasoning CUAをどう組み合わせるのが安全か。

## EVIDENCE TABLE
1) official product docs | "Automate web and desktop apps with computer use" | Microsoft | https://learn.microsoft.com/en-us/microsoft-copilot-studio/computer-use | updated 2026-07-03 / accessed 2026-09-14 | Copilot Studio computer use; OpenAI CUA / Claude Sonnet 4.5/4.6 / Claude Opus 4.6; billing 5 or 15 Copilot Credits per step | locator: Licensing / model table / overview | VERIFIED_SPEC + PRODUCT_VOLATILE | support/context
2) official product docs | "Configure where computer use runs" | Microsoft | https://learn.microsoft.com/en-us/microsoft-copilot-studio/configure-where-computer-use-runs | updated 2026-08-03 / accessed 2026-09-14 | Power Automate for desktop >= 2.61.132.25266 for BYOM; dedicated machine recommendation; sequential queue on same machine | locator: Bring-your-own-machine / Run scheduling and queuing | VERIFIED_SPEC + PRODUCT_VOLATILE | support/limit
3) official product docs | "Run unattended desktop flows" | Microsoft | https://learn.microsoft.com/en-us/power-automate/desktop-flows/run-unattended-desktop-flows | accessed 2026-09-14 | managed Windows user session; target-device session constraints; RDP permission required; authored vs unattended screen-resolution divergence | locator: unattended run behavior and session requirements | VERIFIED_SPEC + PRODUCT_VOLATILE | support/limit
4) official product docs | "Robots" | UiPath Orchestrator | https://docs.uipath.com/orchestrator/automation-cloud/latest/user-guide/about-robots | accessed 2026-09-14 | Automation Cloud current docs; unattended foreground UI requires host-machine credentials; background Session 0 can run without user credentials | locator: With credentials / Without credentials | VERIFIED_SPEC + PRODUCT_VOLATILE | support/context
5) official product docs | "Configuring automation capabilities" | UiPath Orchestrator | https://docs.uipath.com/orchestrator/automation-cloud/latest/user-guide/configuring-automation-capabilities | accessed 2026-09-14 | robot accounts as non-human identities; roles authorize Orchestrator resources; unattended execution bound to infrastructure | locator: Enabling unattended automation on a robot account | VERIFIED_SPEC + PRODUCT_VOLATILE | support
6) official platform spec/docs | "UI Automation Specification" | Microsoft Windows | https://learn.microsoft.com/en-us/windows/win32/winauto/ui-automation-specification | accessed 2026-09-14 | Windows UI Automation element tree exposes controls/properties/patterns from desktop root | locator: UI Automation Elements / UI Automation Tree | VERIFIED_SPEC | context
7) official security docs | "UI Automation Security Overview" | Microsoft | https://learn.microsoft.com/en-us/dotnet/framework/ui-automation/ui-automation-security-overview | updated 2025-06-17 / accessed 2026-09-14 | uiAccess=false by default; protected UI/high-integrity interaction requires explicit trust/signing/manifest conditions | locator: User Account Control / Manifest File | VERIFIED_SPEC | limit/security
8) official failure report | "Fix: Power Automate desktop Windows credential dialog issues" | Microsoft | https://learn.microsoft.com/en-us/troubleshoot/power-platform/power-automate/desktop-flows/windows-credentials-dialog-issues | updated 2026-03-20 / accessed 2026-09-14 | Jan/Feb 2026 Windows security updates caused credential-dialog automation failures; cloud unattended runs cannot use the local-admin workaround | locator: affected KBs / Workarounds | SOURCE_REPORT + PRODUCT_VOLATILE | counter/failure
9) primary research + author-maintained implementation | "OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments" | Xie et al. / xlang-ai | https://github.com/xlang-ai/OSWorld | NeurIPS 2024; repo accessed 2026-09-14 | OSWorld / OSWorld-Verified; real computer tasks across Ubuntu/Windows/macOS | locator: README / public evaluation / citation | RESEARCH + OSS_PRACTICE | benchmark/context
10) primary research implementation | "OS-Blind: The Blind Spot of Agent Safety" | lime-nlp | https://github.com/limenlp/OS_Blind | accessed 2026-09-14 | 300 human-crafted tasks, 12 safety categories; environment-embedded and agent-initiated harms | locator: Dataset Overview | RESEARCH + OSS_PRACTICE | counter/security

## FINDINGS
- Desktop automation has at least two execution paradigms: (A) selector/UIA/action-graph based deterministic RPA and (B) screenshot/vision/reasoning based CUA. They can share host/session/credential infrastructure but should not share the same reliability assumptions.
- Windows UI Automation provides structured control semantics; CUA can act when selectors/APIs are unavailable or UI changes, but model reasoning introduces probabilistic action selection. Future Core should therefore prefer API/structured automation where available, deterministic UI automation next, and CUA as a bounded fallback or explicitly governed execution path rather than treating CUA as a drop-in replacement.
- unattended foreground automation is an identity-and-session problem, not merely a model capability problem. Microsoft and UiPath both require execution infrastructure and user/session credentials or equivalent non-human identity configuration for foreground UI interaction.
- Dedicated execution machines/VMs are operationally meaningful. Microsoft explicitly recommends dedicating a machine for computer use; same-machine jobs queue sequentially. Shared interactive desktops create interruption, state leakage, contention and attribution problems.
- OS privilege boundaries remain effective constraints. Windows UIA cannot simply cross protected/high-integrity UI by default. AI reasoning does not remove UAC/uiAccess boundaries.
- conventional RPA is not immune to platform drift: a 2026 Windows security update broke credential-dialog automation. Therefore deterministic selectors are more reproducible than CUA, but not invariant to OS/security changes.

## LIMITS / COUNTEREVIDENCE / FAILURE ENVELOPE
- Vendor statements that CUA "adapts" to UI changes are product claims, not guarantees of task correctness across arbitrary layouts.
- OSWorld/OSWorld-Verified evaluates bounded benchmark tasks; scores do not establish production SLA, security, or success on organization-specific apps.
- Desktop agents can act correctly at the GUI level yet violate hidden workflow/business rules; application semantics must be separately validated.
- State hazards include modal dialogs, focus theft, resolution/DPI variance, localization, screen locking, concurrent user sessions, stale credentials, pop-ups, remote-desktop differences and application updates.
- Security hazards include credential phishing, data exfiltration, destructive OS actions, malware/code execution, deceptive agreements and prompt/instruction content rendered inside applications. OS-Blind demonstrates these as evaluation categories; it does not prove prevalence in any particular product deployment.
- A robot/service identity with correct infrastructure permissions can still hold excessive application-level permissions. Orchestrator/machine authorization and target-application authorization are separate gates.

## AVAILABILITY_GATES
- Microsoft Copilot Studio computer use: GA announced May 2026; requires generative orchestration; model choice and Copilot Credit billing are current/product-volatile. BYOM requires supported Power Automate for desktop version and machine registration/configuration.
- Microsoft unattended desktop flows: Windows edition/session/RDP/NLA/Entra configuration constraints apply; sovereign cloud/hosted capacity limitations vary by feature.
- UiPath: foreground unattended UI requires appropriate machine login credentials and configured robot/non-human identity; background Session 0 differs materially.
- Protected Windows UI: uiAccess/signing/install-location/UAC conditions can block automation even when ordinary application UI is accessible.

## IMPLEMENTATION_STATE
MATURE for deterministic enterprise RPA; DEPLOYED/GA but rapidly evolving for CUA-assisted desktop automation. Recommended implementation pattern for future book: task router -> API/connector path when available -> deterministic RPA/UIA path -> governed CUA fallback; all paths terminate in common identity/policy/audit controls. Separate attended, unattended-foreground, background/session-0, and hosted/dedicated-machine profiles.

## BENCHMARK_OR_EVAL
Use OSWorld/OSWorld-Verified for cross-application capability evaluation and OS-Blind for safety evaluation. Do not merge capability score with safety score. For organization acceptance, create a private evaluation corpus from real applications with: deterministic completion checks, forbidden-action checks, business-rule assertions, recovery/retry behavior, resolution/localization variants, and clean-state reset between runs.

## SECURITY
least-privilege robot identity; dedicated machine/VM; prohibit unnecessary admin/uiAccess; policy-gate credential/system/security-setting changes; treat on-screen text as untrusted input for CUA.

## PRIVACY
screenshots/video/action traces may capture sensitive desktop content; retention/access/redaction policy required separately from model prompts.
AUDIT: log task request, initiating human/service identity, robot/agent identity, machine/session, model/version when applicable, observations, intended action, executed action, result, approval/override and artifact changes. Screenshot-only replay is insufficient without structured action/result metadata.

## COST
deterministic RPA cost is largely infrastructure/license/run-time; CUA adds per-step model/agent consumption. Microsoft currently documents 5 Copilot Credits per standard computer-use step and 15 for premium models; re-check at publication.

## OPERABILITY
machine image/version pinning, queue depth, stuck-session recovery, credential rotation, selector drift detection, model regression tests and rollback are first-class operations concerns.

## SAFE_WORDING_FOR_FUTURE_BOOK
「Desktop Agent／AI-RPAは、従来RPAをAIで置き換えた単一カテゴリではない。API、structured UI automation、vision/reasoning computer useは異なる失敗様式を持つ実行経路であり、同じidentity・policy・machine isolation・audit境界の下で使い分ける。GUI操作が可能であることは、業務ルール上正しいこと、権限上許可されること、再現可能であることを意味しない。」

## UPDATE_WORK_HANDOFF
1. 公開後本文では「desktop execution hierarchy」をCore化し、API > structured deterministic automation > governed CUAという優先経路を候補にする。
2. provider mappingではMicrosoft Copilot Studio/Power AutomateとUiPathを、機能比較ではなくidentity/session/machine/queue/audit軸でマッピングする。
3. publication時にMicrosoft computer-use GA、model lineup、credit rate、Power Automate minimum version、UiPath current Automation Cloud docsを再確認する。
4. 企業向けTEST_PROFILEを作る場合は専用VM、無害なsynthetic app/data、non-admin test identity、network allowlist、clean snapshot restoreを用い、real credentials/production tenantは使わない。

## OPEN_UNKNOWN
provider横断でCUA action trace schemaを標準化できるか、business-rule violationを一般化して検出する評価標準、各providerのscreenshot/action retentionとtenant-controlled redaction差分、CUAとdeterministic RPAの自動fallback/rollback品質。

## STOP_CONDITION
EA-050初回Packetとして、Core構造、current provider実装例、OS privilege/session failure、benchmark/safety反証、availability/cost gates、後続再確認点が揃ったため停止。次回はEA-051へ進む。

---

## AUDITED_BOUNDARY
Desktop Agent/AI-RPA is layered execution design; API, deterministic UI automation and vision/reasoning CUA have distinct failure modes. Model lists, credits, hosted-machine details and tenant availability remain PRODUCT_VOLATILE. GUI action capability does not establish business-rule correctness, authorization or reproducibility. Preserve all OPEN_UNKNOWN.

NEXT: /agent/start.json
STOP: This is a knowledge projection, not an execution order.