AnswersPublished19 sources
InAgent’s 90.2% OSWorld Score and the Rise of Harness Engineering
Intelligence Indeed reported that its InAgent/Z Agent reached 90.2% on the 361 task OSWorld configuration, the first reported score above 90%. OSWorld evaluates agents in real desktop environments with execution based checks, not just text responses.
Published byEdited with GPT-5.6 TerraImages generated with GPT Image 2
Research answer
Key takeaways
- Intelligence Indeed reported that its InAgent/Z Agent reached 90.2% on the 361 task OSWorld configuration, the first reported score above 90%.
- OSWorld evaluates agents in real desktop environments with execution based checks, not just text responses.
- A 90.2% pass rate would still leave 9.8% of comparable tasks incomplete, so high impact workflows need deterministic checks, permissions, exception queues, and human approval paths.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: How did Chinese company Intelligence Indeed’s InAgent become the first computer-use agent to exceed 90% on the OSWorld benchmark—achieving 9. Article summary: The reported result is best understood as a system-level achievement, not evidence that InAgent has a uniquely superior foundation model. InAgent reportedly combined model reasoning with an execution “harness” that plans. Topic tags: general, academic, general web, user generated. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, char
Intelligence Indeed reported in August 2026 that its computer-use product—called Z-Agent in its announcement and InAgent in other coverage—achieved a 90.2% task-success rate on OSWorld and became the first reported participant above the 90% mark. The company said the score also placed it first on an agentic-framework sub-leaderboard. 18
22
The important lesson is not that a single model has solved desktop automation. It is that agent performance increasingly depends on the harness around the model: the software that turns a model’s plan into controlled actions, checks whether those actions worked, and handles the many ways a real workflow can fail.
What InAgent reportedly achieved
Coverage of the result attributes these figures to InAgent’s OSWorld evaluation:
- 90.2% overall task success
- 84.7% on cross-application tasks
- 92.3% on GIMP tasks
- 100% on system-level tasks
22
These are reported results, not independently audited findings. The supplied reporting does not include a public technical report that isolates the foundation model, action tools, prompts, evaluation settings, or harness components responsible for each score. That distinction matters: an agent benchmark score measures a model-and-system configuration, not model intelligence in isolation.
It is also worth separating a historical claim from a permanent leaderboard position. A September 2026 third-party leaderboard snapshot lists a different system, Qwen3.8 Max, at 86.1% on OSWorld-Verified, illustrating how leaderboard standings and evaluation configurations can change over time. 3
Why OSWorld is a meaningful test
OSWorld was designed to test agents operating real computer environments rather than answering questions about screenshots or simulated webpages. The original benchmark includes 369 real-world tasks across applications such as Chromium, LibreOffice, GIMP, Thunderbird, VLC, and Visual Studio Code, plus operating-system and multi-application work. Tasks are evaluated from configured environment states using execution-based checks. 2
14
The commonly used 361-task configuration excludes eight Google Drive tasks that may require manual setup. 14
That makes OSWorld more demanding than a simple click-prediction test. A successful agent must navigate interfaces, maintain progress across steps, manipulate files or application state, and leave the machine in the expected final condition.
The historical context makes a 90.2% claim notable. OSWorld’s original published reference reported humans completing more than 72.36% of tasks, while the best model achieved 12.24%. 14 But the comparison should not be treated as a universal measure of workplace competence: benchmark performance depends on task selection, environment setup, action budgets, tools, and the agent harness.
The shift from prompts to harnesses
A prompt tells a model what to do. Context engineering determines what information it receives. Harness engineering designs the runtime system that controls how the agent observes, acts, persists state, verifies completion, and recovers from failures.
A survey of agent-system design describes the execution harness as infrastructure that shapes an agent’s perception, actions, and ability to detect and recover from errors. It includes elements such as observation design, action spaces, execution sandboxing, context management, and verification loops. 28
For a computer-use agent, a capable harness commonly needs to do six jobs:
- Break a workflow into testable subgoals. Instead of asking an agent to “complete an order update,” the system can require it to locate the record, validate fields, make the update, attach evidence, and verify that the change persisted.
- Route actions to the right interface. An API, command, or structured tool can be preferable to a mouse click when it is available and permitted; GUI control can handle steps that lack a machine-readable interface.
- Maintain durable state. The agent needs a record of what it has completed, the current application state, selected entities, files, and whether it is safe to retry an action.
- Verify outcomes from the system of record. A model’s statement that it succeeded is not proof. A harness can check an API response, database value, saved file, or application state.
- Recover safely. It needs policies for stale pages, failed clicks, pop-ups, timeouts, partial submissions, and authentication interruptions.
- Enforce permissions and observability. Logs, action traces, approval gates, and constrained credentials make consequential work reviewable and safer.
There is empirical evidence that this layer can materially affect agent performance. One study reported that ten iterations of agentic harness engineering raised Terminal-Bench 2 pass@1 from 69.7% to 77.0%, exceeding the cited 71.9% human-designed harness baseline. That is a different benchmark, so it does not validate InAgent’s OSWorld score—but it supports the broader claim that runtime design can produce substantial gains without simply swapping in a new model. 29
Why a hybrid API-and-GUI agent is commercially attractive
The reported strategic appeal of InAgent is a hybrid approach: use reliable programmatic interfaces where they exist, while retaining human-like GUI control for applications that do not offer suitable APIs. 22
That design addresses a practical enterprise problem. Many workflows span modern services with usable APIs and older desktop software, vendor portals, or internal systems where the only practical path is the screen a human employee uses.
In principle, the agent can take the most reliable path at each step:
- Use an API or command for structured retrieval and updates.
- Use GUI interaction for a legacy or closed interface.
- Check the outcome through a deterministic system of record when possible.
Research on OSWorld-MCP similarly finds that tool invocation can improve computer-use task success in some settings. In the reported experiments, adding MCP tools improved OpenAI o3 from 8.3% to 17.6% at 15 steps and Claude 4 Sonnet from 38.9% to 45.0% at 50 steps. 24
The caveat is that GUI automation remains exposed to interface changes, latency, pop-ups, ambiguous targets, access controls, MFA, and irreversible actions. Hybrid execution can reduce those risks, but it does not eliminate them.
What 90.2% means for deployment
A 90.2% completion rate implies that 9.8% of benchmark tasks did not complete successfully. On a hypothetical 10,000 comparable tasks, that is about 980 exceptions before accounting for errors that appear successful but produce an incorrect business result.
That does not make the score unimportant. It suggests that computer-use agents may be approaching usefulness for bounded, repeatable workflows with strong controls. But it is not a license for unattended automation of payments, legal filings, safety-critical operations, or customer-impacting changes.
For production use, teams should evaluate more than a headline pass rate:
- completion and correctness on their own workflows;
- failure modes and recoverability;
- cost, latency, and action volume;
- permission boundaries and approval points;
- evidence that each completed action actually changed the intended system state;
- the size and handling process for the exception queue.
The newest, harder OSWorld 2.0 benchmark reinforces the need for caution. It contains 108 long-horizon workflows, and its reported top binary-completion result was 20.6% under its primary 500-step metric—evidence that agents remain far from broadly professional-grade computer use when tasks become longer and more realistic. 17
The real milestone: execution engineering
InAgent’s reported 90.2% result is best read as a milestone for agent-system engineering. Better prompts and richer context still matter, but reliable automation increasingly comes from the machinery around the model: decomposition, tool selection, persistent state, verifiable completion, bounded authority, and recovery paths.
For enterprises, the useful question is not simply which model tops a benchmark. It is whether an agent can execute a specific workflow with measurable correctness, controlled side effects, and a safe handoff when it cannot proceed. That is the standard a high OSWorld score can help inform—but cannot, by itself, prove.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "InAgent’s 90.2% OSWorld Score and the Rise of Harness Engineering"?
Intelligence Indeed reported that its InAgent/Z Agent reached 90.2% on the 361 task OSWorld configuration, the first reported score above 90%.
What are the key points to validate first?
Intelligence Indeed reported that its InAgent/Z Agent reached 90.2% on the 361 task OSWorld configuration, the first reported score above 90%. OSWorld evaluates agents in real desktop environments with execution based checks, not just text responses.
What should I do next in practice?
A 90.2% pass rate would still leave 9.8% of comparable tasks incomplete, so high impact workflows need deterministic checks, permissions, exception queues, and human approval paths.
Sources
globenewswire.comIntelligence Indeed's Z-Agent Becomes OSWorld's First 90%+
benchlm.aiOSWorld-Verified Leaderboard & Scores — August 2026
arxiv.orgOSWorld-Human: Benchmarking the Efficiency of Computer-Use ...
osworld-v1.xlang.aiOSWorld: Benchmarking Multimodal Agents for Open-Ended ...