# Desktop agent > VI: https://5ac.vn/wiki/desktop-agent/ · EN: https://5ac.vn/en/wiki/desktop-agent/ · lang: en · spec_as_of 2026-10-19 (spec no formal spec) · cluster C7 An AI agent running on the employee's own machine: it takes work from a queue, operates email, spreadsheets and CRM, and returns logged results — it processes tasks, not questions. --- ## Definition A desktop agent is an AI agent that runs on the employee's work machine — or on the company's own server — takes work from a queue, operates business tools such as email, spreadsheets, CRM and the browser, and returns results with a log. The line against chatbots sits in task processing: a chatbot answers questions, while a desktop agent pulls the follow-up list, checks the CRM, prepares the draft, assigns an owner and records the action. 5ac's 24/06/2026 piece sets three conditions for deployment: the task repeats often, the output can be inspected, and mistakes are not catastrophic once a human approves the final step. ## Mechanism The research lineage is the agent–computer interface: SWE-agent (arXiv:2405.15793) shows that how the interface is built — which file the agent opens, which command it runs, what it sees on screen — decides how well it operates a computer, and that work opened the coding-agent branch. Desktop agents carry the same interface question out of the code editor and onto the daily work surface: email, spreadsheets, CRM screens, internal chat windows. Four layers of a running desktop agent: | Layer | What it does | What to watch | | --- | --- | --- | | Entry point | present on the familiar work machine, calls workflows from an office-style interface instead of the command line | frontline staff can use it without quietly widening their own rights | | Execution loop | takes work from a queue, moves across several application windows, returns logged results | one task can touch email, spreadsheets and CRM at the same time | | Permissions | rights by role, a secret vault, logging switched on by default | the agent must not read email, edit the CRM and send financial files in one go | | Approval | a stop point for high-risk actions | who approves, when, and whether the decision is recorded | Three deployment models to compare before choosing: an agent inside a closed platform (fast trials, weak cost and data control), an agent through cheap APIs (fits high-volume repetitive work, needs good workflow design), and a desktop agent on a private VPS (fits workflows that touch internal data, requires operating discipline). ## Application context A desktop agent sits between two familiar options: legacy RPA copies screen actions but breaks when the interface changes, while SaaS automation starts fast but fragments data. The real gain is onboarding — a founder does not have to teach the whole team the command line — and the Windows Native piece names three risks to lock early: over-broad permissions, no audit trail, and control arriving after integration. The larger warning comes from the control-plane piece: as agents move from terminal to desktop their reach expands sharply, so desktop agents turn governance into a product requirement rather than a late IT task. And because the desktop is only the interface layer, 5ac marks its limits: orchestration, permissions, audit logs and private hosting are what actually hold a company together. The first workflow of a small business should save several hours a week and never make financial or legal decisions on its own. ## Sources fetched - [Desktop agents for business automation: where should Vietnamese SMBs start?](/en/blog/business-automation-desktop-agents-vietnam-smb/) — three deployment models, three conditions, four safeguards — accessed 2026-10-19 - [Windows-Native Desktop AI Agents: The New Automation Layer for Small Businesses](/en/blog/desktop-ai-agents-windows-native-small-business-automation/) — legacy RPA versus SaaS automation, three workflows for the first 30 days, three risks — accessed 2026-10-19 - [Desktop AI Agents for Business Automation: From CLI Tools to SMB Operating Systems](/en/blog/desktop-ai-agents-business-automation/) — five layers deeper than the interface, CLI and desktop as two entry points — accessed 2026-10-19 - [The Agent Control Plane War Has Begun — Why Vietnamese SMBs Should Choose Open Source Over AI Super Apps](/en/blog/agent-control-plane-war-vietnam-smb/) — desktop agents turn governance into a product requirement — accessed 2026-10-19 - — SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering (v3) — verified via arXiv API 2026-10-19 ## Related - [AI Agent](/en/wiki/ai-agent/) - [Coding agent](/en/wiki/coding-agent/) - [Business automation](/en/wiki/business-automation/) - [Agent orchestration](/en/wiki/agent-orchestration/) - [Multi-agent OS](/en/wiki/multi-agent-os/)