Desktop agents: the most-used
features are worth the least

The desktop agent features enterprises use most — document drafting at 51.8% and spreadsheet work at 38.2% — are the ones with the least long-term value. Cross-system data coordination, which B2B buyers widely rank as the highest long-term value, sits at 14.3% usage. This is not an adoption problem, it is an evaluation problem: usage measures low friction, not value. Our judgment: 14.3% reflects a harness gap — permissions, evidence and orchestration are not yet trustworthy enough to cross systems.

1. A Chinese-market report finally treats model plus harness as the unit

Frost & Sullivan, with LeadLeo Research Institute, published the 2026 Global Desktop AI Agent Market Research Report on 20 September 2026. Its most useful content is not the market sizing but the framing of the harness as the unit of evaluation: the harness is the whole system outside the model — system prompt, tool calling, filesystem, sandbox, orchestration logic, hook middleware, feedback loops, constraint mechanisms. The model only reasons and generates; the harness connects state, tools, feedback, execution environment and security boundaries so the agent can actually work. Source: the report excerpt page.

That section is labelled as a literature review compiled by the analyst, and relays Anthropic’s definition: the harness is operating-system-level infrastructure outside the model, managing context and memory, scheduling tools, collecting execution results and maintaining permission boundaries. The report’s conclusion follows: evaluating an agent’s real capability has never been about the model itself but about the model-plus-harness combination — the same model under a different harness can perform very differently.

Citation discipline: this is a commercial research product, the public page is an excerpt, and neither sample size nor methodology is disclosed. The usage figures here can only be cited as the report’s framing, not as independent measurement or audit evidence. The harness section is itself a literature review, not original measurement.

2. The usage distribution exposes friction, not demand

The report’s enterprise usage distribution is: document drafting 51.8%, spreadsheet data processing 38.2%, cross-system data coordination 14.3%. The ordering is tidy — the lower the friction, the higher the usage. Drafting needs almost no system permission; spreadsheet work needs only file access; cross-system coordination must simultaneously handle authorisation, credentials, state consistency and audit trails.

The report names the mismatch itself: although cross-system data coordination currently accounts for 14.3%, B2B users widely treat it as the most valuable long-term differentiator; deep industry research and cross-system coordination involve complex judgement, so automation is low today but growth potential is the largest.

Our judgment: usage is a proxy for low friction, not for value. Using it as a deployment priority steers budget into personal-productivity features — which happen to be the features every competitor ships by default, differentiating nothing. The real question is why cross-system automation is so low. Not because users do not want it, but because the permission model and evidence chain for cross-system execution do not exist yet.

3. Single-vendor telemetry cannot see the cross-system range

There is a structural blind spot here: a desktop agent vendor’s telemetry only covers the segment it participates in. One cross-system task leaves three disconnected records in three systems, and no party sees the complete execution trace. Usage statistics therefore systematically undercount cross-system tasks — they cannot count what they cannot see.

That also explains how low usage and highest value can both be true: value happens between system boundaries, while measurement happens inside them.

Our judgment: runtime evidence of cross-system execution — who authorised what, which system was called, how the result was verified — is an uncommodified capability. It is neither model capability nor interface capability; it is part of the harness. And it sits exactly in the range single-vendor telemetry structurally cannot reach.

4. The real definition is a persistent orchestration layer

The report defines a desktop AI agent as an AI assistant running on the user’s desktop with autonomous execution capability, in substance a persistent orchestration layer embedded in the desktop environment, perceiving screen state, filesystem and application context in real time.

Note “persistent orchestration layer”: it means the competition is not in single-turn conversation quality but in long-lived state, permission boundaries and recoverability. A desktop agent that runs one task is a demo; one that holds state across sessions and can roll back on failure is a production component.

The report’s segmentation points the same way: data deployment (local, cloud SaaS, hybrid, private enterprise) times vertical (general office, engineering IDE, finance data, design and creative) times product form (standalone client, embedded in office suites, open-source framework, enterprise platform), with two parallel paths — native agents driving the system directly, and agents embedded in existing office workflows. For enterprise buyers, selection is a three-way crossing, not a question of which model is smarter.

5. Our judgment and a deployment checklist

Our judgment: desktop agent deployment priorities should be ranked by value times automation difficulty, not by usage. Usage walks you to the easiest ground; value sits in the opposite direction.

Action: pick one process your team repeats hundreds of times a month that spans three systems, get it running under a minimal harness with recorded execution evidence, and only then discuss scale. The question to leave with: does your desktop agent selection criteria include a line for verifiable cross-system execution? If not, what you bought is probably a personal productivity tool.

References: Frost & Sullivan with LeadLeo Research Institute, 2026 Global Desktop AI Agent Market Research Report, 20 September 2026. Commercial research product; the public page is an excerpt with no disclosed sample size or methodology, so all usage figures cited here are that report’s framing, not independent measurement. frostchina.com report excerpt. The harness section is a literature review relaying Anthropic’s definition.