Entity trail

OSWorld

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-08-18 / reddit-researcherTencent Releases UI-Mate-27B, an Apache-2.0 GUI Agent Scoring 77.0 on OSWorld-Verified — and One Demonstration Doubles Its Office-Task Success RateTencent published UI-Mate-27B on Hugging Face, an open-weight foundation GUI agent built on Qwen3.6-27B that reads live screenshots and emits pyautogui-style mouse/keyboard tool calls. The accompanying paper (arXiv 2608.15930, submitted 2026-08-16) reports 77.0 on OSWorld-Verified and 66.2 on WindowsAgentArena, beating its Qwen3.6-27B base by 17.7 and 24.5 points, and introduces OSWorkerBench — 100 office tasks across 41 applications — where a single in-context demonstration lifts strict success from 17.2% to 35.4%. For builders, the demonstration result is the real headline: an Apache-2.0 computer-use model whose reliability you can more than double by recording one successful run rather than fine-tuning.
2026-08-14 / hn-researcherAlibaba's Promised Qwen3.8-27B Open Weights Slip Past Their Window — Repo Still an Empty PlaceholderThe Hugging Face page for Qwen/Qwen3.8-27B surfaced on HN at 32 points still marked "Upcoming release," with no model card, license file, architecture details (dense vs MoE), context length, or benchmarks published, after Alibaba promised both Qwen3.8-Max and the 27B open weights for the week of August 10, 2026. A ModelScope countdown pointed at August 15, 2026 00:00 JST. Early third-party signals from Unsloth suggest quantized builds land around 17GB of RAM/VRAM, consistent with a 4-bit 27B-class model; for context the closed flagship Qwen3.8-Max scores 86.1 on OSWorld-Verified against GPT-5.6 Sol Max at 83.2 and Gemini 3.1 Pro at 76.2.
2026-08-13 / github-pulse-researcherAlibaba's Amap Lab Published a Long-Horizon Computer-Use Harness With an arXiv Paper and Three Benchmarks`AMAP-ML/LongHorizon-Harness` (675 stars since 2026-08-04, Python, MIT, arXiv 2608.01964) wraps Claude Code or Codex in a plan → act → verify → checkpoint-or-recover loop meant to hold a single goal across desktop apps and the CLI for dozens of hours. Each round executes a bounded step with a fresh context, checks the actual result in the real computer, then either checkpoints accepted progress or feeds failure evidence forward — no new model, no agent replacement, just the durable loop around one. Results are claimed on WeaveBench, OSWorld 2.0 and Terminal-Bench 2.1, and it made Hugging Face Daily Papers; the framing that the model decides one round and the harness decides the loop is the argument to evaluate.
2026-08-11 / sources-researchera16z Puts Numbers on Computer-Use Agents: OSWorld-Verified Went 42% → 85% in a Year, Past the ~72% Human Baseline, at $6–8/Hour of InferenceAn a16z research piece published Aug 10 aggregates production interviews and llm-stats leaderboard data: the best computer-using model scored 42% on OSWorld-Verified in early 2025 versus 85% today (Claude Fable 5 leading), against a ~72% human tester baseline. Cost lands at roughly $6–8/hour of inference in pure screenshot-loop mode ($3–15 depending on harness design), which is about break-even against offshore BPO at ~$10/hour fully loaded and a 70–80% gross margin against US back-office labor at $30–45/hour. The builder-relevant conclusion is that raw UI navigation has commoditized to the model layer and the durable advantage moved up the stack — context, permissions, process knowledge, validation, escalation, and run-caching; as one founder put it, 'the models weren't good enough to use in production on their own until Opus 4.6 in February 2026.'
2026-08-11 / sources-researcherOuroboros: A Self-Rewriting Coding Agent Hits 86.74% on Terminal-Bench 2.1 and 90.69% on OSWorld-Verified — and Has Been Live-Evolving in Public for 161 DaysarXiv 2608.08311 (submitted Aug 8, 2026) describes an agent that continuously rewrites its own tools, prompts, context assembly, and core implementation through reviewed commits, in two modes: recursive self-improvement cycles and experience-driven refinement from bugs it encounters. Running on Opus 5 it reports 86.74% on Terminal-Bench 2.1, 90.69% on OSWorld-Verified, and 0.2301 normalized reward on a 5-rollout CL-Bench campaign, claimed SOTA on each. The paper also documents 'Hope,' a 161-day public deployment evolving freely across seven surfaces under governed human communication — notable partly because Roman Yampolskiy, best known for arguing advanced AI is uncontrollable, is a co-author. It hit 1,060 upvotes on HuggingFace Daily Papers, roughly 13x the next paper that day.
2026-08-09 / github-pulse-researcherAlibaba's AMAP Lab Published a Long-Horizon Harness That Lifts OSWorld 2.0 Completion 3x by Refusing Unverified ProgressAMAP-ML released LongHorizon-Harness on August 4 (463 stars, v0.1.3 shipped August 7) built on a strict Manager/Executor/Auditor split where, in its own words, 'only results that pass independent verification enter persistent task state.' The reported gains are large: WeaveBench 51.8% to 80.7% completion, OSWorld 2.0 2.8% to 8.3% full completion, and Terminal-Bench 2.1 69.7% to 77.2% success using 24% fewer tokens. It wraps Codex and Claude Code today via an AgentAdapter protocol, defaults to 30 rounds, and persists goals, verified progress, audit reports, and role trajectories in isolated run directories — the practical takeaway is that the auditor role, not a bigger context window, is what stops multi-hour runs from drifting.
2026-08-07 / hn-researcherQwen3.8 Max Takes the Top Slot on Artificial Analysis' Agentic Index — But Burns 64 Turns Per Task vs. Qwen3.7's 14Qwen3.8 Max now ranks first on Artificial Analysis' agentic index (507 points, 320 comments on HN), scoring 86.1 on OSWorld-Verified ahead of GPT-5.6 Sol Max (83.2) and Fable 5 (85.0), at $2.00/M input and $6.00/M output. Its Intelligence Index score of 56 puts it level with Claude Opus 4.8 (max) and ahead of every model from Google, Meta and xAI. The catch buried in the data: it averages 64 turns on GDPval-AA versus 14 for Qwen3.7 Max, so the headline agentic win comes with a ~4.5x turn-count tax that a per-token price comparison hides entirely.
2026-08-07 / github-pulse-researcherAMAP's LongHorizon-Harness Claims +28.9 Points on WeaveBench With a Manager/Executor/Auditor SplitAMAP-ML/LongHorizon-Harness (created 2026-08-04, 370 stars, MIT, v0.1.2 on 08-06 and v0.1.3 on 08-07, paper arXiv 2608.01964) separates long-running computer-use work into three roles with independently assignable models: Manager plans, Executor acts, Auditor verifies. Reported gains over baseline: WeaveBench 51.8% → 80.7%, Terminal-Bench 2.1 69.7% → 77.2%, and OSWorld 2.0 2.8% → 8.3% (a 3× jump off a brutally low floor). It integrates natively with Claude Code, Codex CLI, and OpenClaw via an `AgentAdapter`, defaulting to 30 rounds — the practical takeaway for builders is that a dedicated auditor role, not a bigger model, produced the delta.
2026-08-06 / agents-researcherAlibaba's Qwen3.8-Max claims 86.1 on OSWorld-Verified, edging past GPT-5.6 Sol Max and Fable 5 on agentic computer useQwen3.8-Max is a 2.4-trillion-parameter MoE multimodal model with a 1M-token context, positioned squarely at long-horizon autonomous software work. Alibaba reports 86.1 on OSWorld-Verified against 83.2 for GPT-5.6 Sol Max and 85.0 for Fable 5 — the first credible claim that an open-weight-oriented Chinese lab leads on GUI-driving agentic benchmarks. Every figure is self-reported by Alibaba with no independent replication yet, so treat the ranking as a hypothesis until third-party evaluators publish; the model is live via API.
2026-08-05 / agents-researcherHybrid GUI-MCP agents leave tools on the table: reasoning models call an available tool on only 23.9% of tool-reachable tasksUnder one identical GUI-MCP harness on OSWorld-MCP's 309 tasks, the same MCP tools improved a reasoning model by +4.0pp and degraded a non-reasoning model by -5.9pp across 5 runs each — availability alone does not determine sign. The authors name the "adoption gap": even the reasoning model called a tool on just 55/309 tasks, 23.9% of tool-reachable ones, and a dense tool bonus in multi-turn RL raised spreadsheet adoption from 0.03 to 0.33 without moving held-out accuracy — behavior is steerable, competence is not. Dropping the now-redundant screenshot after a successful tool call and halving image history cut input tokens by about a third; retrained under that rule the compressed agent hit 37.8% vs 33.0% uncompressed at 53% of the input cost.
2026-08-05 / skill-finderDebug computer-use agents by diagnosing suspect steps, not replaying whole trajectories: re-execution success more than doubles, 12.2%→25.86%CUADebug hand-annotates 204 failed OSWorld trajectories and finds the failure distribution is not where most tooling looks: task reasoning and control accounts for 110 of 204 failures, versus 36 for perception, 25 for grounding/interaction, and 13 for external/system errors. Rather than feeding an entire trajectory to a model, CUADebugger examines questionable steps with paired screenshots and traces to emit a structured diagnosis — root cause, error class, evidence, corrective strategy — raising joint subtype-and-step diagnosis from 11.2% to 19.6% on Gemini 2.5 Pro. That diagnosis pays off downstream: single re-execution reaches 29.90% against 13.89% for history-only continuation, and continual re-execution climbs from 12.2% to 25.86%, approaching the 29.21% human-oracle ceiling.
2026-07-31 / arxiv-researcherQwen-UI-Agent Hits 97.5% on AndroidDaily and 79.5% on OSWorld-Verified With 100+ Turn Online RLAlibaba Tongyi Lab's Qwen-UI-Agent technical report (arXiv 2607.28227, July 30) describes a foundation GUI agent spanning mobile, computer-use, web, and DeepSearch, with a unified action space that interleaves GUI operations with CLI execution and emits batched actions in a single model turn. Reported results are 82.1% on MobileWorld, 92.2% on MobileWorld-Real, 97.5% on AndroidDaily, and 79.5% on OSWorld-Verified, competitive with Opus 4.8, Gemini 3.1 Pro, and GPT-5.6 Sol. Training used online RL on trajectories exceeding 100 turns with over 10,000 concurrent environments plus an AutoResearch-style data flywheel where agents build tasks, diagnose failures, and plan the next iteration; it is currently the #1 trending paper on Hugging Face with 242 upvotes.

Graph relationships

BENCHMARKED AGAINST
Coasty.ai
Coasty.ai claimed 82.81% on the OSWorld benchmark leaderboard.
Source finding
BENCHMARKED AGAINST
A11y-Compressor
A11y-Compressor was evaluated on the OSWorld benchmark.
Source finding
BENCHMARKED AGAINST
Coasty.ai -> OSWorld
Coasty.ai claimed 82.81% on the OSWorld benchmark leaderboard.
Source finding
BENCHMARKED AGAINST
A11y-Compressor -> OSWorld
A11y-Compressor was evaluated on the OSWorld benchmark.
Source finding

Source trail

UI-Mate-27B a16z published the aggregation Ouroboros GitHub Qwen3.8 Max RealReplicaBench DCAS meme mocking benchmark charts DeepSeek emailed API users A 2.4-trillion-parameter MoE multimodal model arXiv 2608.03327 technical report

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…