Assistant Benchmark
1–10 mean of scored dimensions · Assistant Benchmark (independent) · US
Updated 8 Oct 2026. 12 products scored; coverage differs by product.Full leaderboard →
Personal Agent Evaluation · Global evidence hub
Every personal AI agent, compared on evidence.
PAEval compiles independent evaluations, hands-on tests and official documentation for personal agents worldwide in one traceable place. Every number links to the source that measured it.
45products tracked · 30 fully profiled
4independent score sources
539source-backed capability records
344cited sources · EN ZH JA FR DE
Personal agents
All tracked products with every independent score available today, each on its source's own scale.
in development
PAEval Index: in development. We are building our own PAEval Index for personal agents. No PAEval score or ranking is published yet; it will appear here, together with its full method, once it is ready. Until then, this page shows independent third-party results on their own scales, alongside documented capabilities.
| # | 1–10 | % | 0–100 | % | ||
|---|---|---|---|---|---|---|
| Grok BotxAI (via Cursor) · United States | 7.3 | 30% | — | 50.0% | 2of 35 records | |
| InstinctInstinct (Spear Street Technology) · United States | 8.5 | 40% | — | 60.5% | 4of 16 records | |
| Meta MuseMeta · United States | 9.3 | 20% | — | 28.9% | 1of 31 records | |
| Gemini SparkGoogle · United States | Listed, unscored | 60% | — | 25.7% | 2of 22 records | |
| OpenAI dotsOpenAI · United States | 8.4 | — | — | 55.9% | 1of 32 records | |
| ArkClawByteDance · China | — | — | 92.61 | — | 0of 0 records | |
| AsaplyAsaply Inc. · United States | 7.5 | — | — | — | 2of 17 records | |
| AutoClawZhipu AI · China | — | — | 93.23 | — | 0of 0 records | |
| CaddyCaddy Technologies Inc. · United States | 7.3 | — | — | — | 2of 16 records | |
| Cue by ManusManus · Singapore | — | — | — | 19.1% | 2of 19 records | |
| DuMateBaidu · China | — | — | 93.44 | — | 0of 0 records | |
| Hark ProHark · Not verified | — | — | — | 44.1% | 0of 0 records | |
| KimiClawMoonshot AI · China | — | — | 96.01 | — | 0of 0 records | |
| MaxClawMiniMax · China | — | — | 94.79 | — | 0of 0 records | |
| MiMoClawXiaomi · China | — | — | 96.74 | — | 0of 0 records | |
| OllieConfabulation Corporation · United States | 8.3 | — | — | — | 3of 17 records | |
| PallyPally Technologies · United States | 8.3 | — | — | — | 2of 17 records | |
| QwenPawAlibaba · China | — | — | 92.97 | — | 0of 0 records | |
| ShuffleDNFT · United States | 7.8 | — | — | — | 2of 15 records | |
| StepClawStepFun · China | — | — | 91.31 | — | 0of 0 records | |
| sznNomadic Futures · United States | 8.2 | — | — | — | 3of 16 records | |
| TomoMapo Labs · United States | 7.7 | — | — | — | 2of 15 records | |
| TownTown.com · United States | 8.3 | — | — | — | 2of 16 records | |
| Wenxin Assistant (task mode)Baidu · China | — | — | 97.62 | — | 0of 0 records | |
| WorkBuddy (Tencent)Tencent · China | — | — | 96.61 | — | 0of 16 records | |
| ArkClaw-LiteByteDance · China | — | — | Mar batch only | — | 0of 0 records | |
| ArkClaw-ProByteDance · China | — | — | Mar batch only | — | 0of 0 records | |
| AutoGLM (hosted application)Zhipu / Z.ai · China | — | — | — | — | 2of 15 records | |
| Claude (including Cowork execution)Anthropic · United States | — | — | — | — | 2of 24 records | |
| CoPawAlibaba · China | — | — | Mar batch only | — | 0of 0 records | |
| Coze 3.0 / Coze Agent (formerly Coze Space)ByteDance · China | — | — | — | — | 0of 14 records | |
| Doubao Phone Assistant (consumer version)ByteDance · China | — | — | — | — | 2of 14 records | |
| DuClawBaidu · China | — | — | Mar batch only | — | 0of 0 records | |
| Hermes AgentNous Research · Open source / community | — | — | — | — | 0of 15 records | |
| HONOR YOYO AgentHonor · China | — | — | — | — | 0of 13 records | |
| Lindy(Personal Lindy + Teammate)Lindy · United States | — | — | — | — | 1of 22 records | |
| Manus 2.0 / Manus StudioManus · Singapore | — | — | — | — | 2of 29 records | |
| Microsoft Copilot Cowork for personal accounts(Preview)Microsoft · United States | — | — | — | — | 2of 20 records | |
| MiniMax Agent/MiniMax Code Work ModeMiniMax · China | — | — | — | — | 1of 12 records | |
| OpenClawOpenClaw Foundation · Open source / community | — | — | — | — | 1of 16 records | |
| Perplexity Computer (Cloud)Perplexity · United States | — | — | — | — | 2of 21 records | |
| Perplexity Personal Computer (local access)Perplexity · United States | — | — | — | — | 2of 20 records | |
| QClawTencent · China | — | — | Mar batch only | — | 0of 0 records | |
| Qwen App (task execution / work assistant)Alibaba · China | — | — | — | — | 1of 14 records | |
| QwenWorkAlibaba / DingTalk · China | — | — | — | — | 0of 14 records |
Default order: number of independent scored sources, then alphabetical — not a quality ranking. Each source column uses its own scale and is never blended with the others.Which sources are scored →
1
Native scales, never blendedEach third-party score stays on its publisher's scale with its sample, date and coverage beside it.
2
Features are not performanceWhat a vendor documents and what an evaluator measured are shown side by side, never merged.
3
Unknown is not unsupportedMissing evidence stays visibly missing. Nothing is scored as zero by absence.
Independent leaderboards
4 evaluators of finished products, using real services or simulated accounts. Each publishes a different scale, sample and product set, so read each chart on its own.
micro1 PersonalAgentBench
% of 10 workflows with trusted completion · whiskers = Wilson 95% interval
n = 10 per agent. Every pair of intervals overlaps.Full leaderboard →
SuperCLUE-XClaw
0–100 weighted total of 5 dimensions · Aug 2026 batch · CN
Products within 1 point share a publisher tier.Full leaderboard →
PersonalAgentEval
MatrAIx · simulated personal accounts · publisher 95% bootstrap intervals
152 scoring units = 109 single tasks + 43 sessions from 23 long-horizon task chains (132 task chains).Full leaderboard →
Capability coverage
What vendors document and third parties observed, by capability domain. A filled cell is evidence of availability, not of task success.
| Product | Autonomy & background work | Execution environments | Calls, messages & identity | Connectors & tools | Memory & personalization | Channels & interfaces | Research & deliverables | Transactions & bookings | Collaboration & parallelism | Reliability & recovery | Permissions, security & data |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Meta Muse | |||||||||||
| OpenAI dots | |||||||||||
| Grok Bot | |||||||||||
| Claude (including Cowork execution) | |||||||||||
| Manus 2.0 / Manus Studio | |||||||||||
| Gemini Spark | |||||||||||
| Perplexity Computer (Cloud) | |||||||||||
| Microsoft Copilot Cowork for personal accounts(Preview) | |||||||||||
| OpenClaw | |||||||||||
| Qwen App (task execution / work assistant) | |||||||||||
| WorkBuddy (Tencent) |
Independently observedDocumentedPartly disclosed/testedLimited / conditionalPlannedHistorical / unconfirmedNot supportedNot disclosedUnknownNo record yetIncludes independent evidence
Head-to-head
Popular comparisons
By capability
Find agents that…
that work in the background with cloud or local computers that make calls and send messages with app connectors and tools with long-term memory in chat apps, voice and other channels for research and deliverables that book and buy for teams and parallel work reliability and recovery permissions, privacy and security self-hosted / open source
Report · 10 Oct 2026
What independent tests say about personal AI agents (October 2026)
Independent evidence remains uneven across products, sources and tested configurations.
Latest update · 28 Oct 2026
PersonalAgentEval added as a fourth scored source
- Imported seven PersonalAgentEval product results from MatrAIx on the publisher’s native resolved-rate and partial-credit scales, with publisher bootstrap intervals.
- Disclosed 109 single tasks plus 43 sessions from 23 long-horizon task chains: 152 scoring units and 132 task chains, with interface and memory-reset caveats.
- Added Hark Pro as a scored-only product; expanded the directory to 45 products without inventing capability facts or a PAEval composite.
What PAEval is not
No vendor-reported rankings
Vendor-reported benchmark results are labelled and kept out of scored columns. Model-level benchmarks (a model inside a harness) are listed separately from product-level evaluations. Editorial feature comparisons are cross-checks, not scores.
Frequently asked questions
What is a personal AI agent?
A personal AI agent acts for one person over time rather than only answering questions: it can carry out tasks across apps, accounts and devices, remember context, and keep working in the background. PAEval covers persistent assistants, delegated work agents, device-integrated agents and self-hosted agents; chat-only assistants and developer coding agents are out of scope unless they offer personal delegation.
Which personal AI agents score highest in independent evaluations?
It depends on the evaluator, and their scores cannot be combined. On Assistant Benchmark (1–10 scale, updated 8 Oct 2026), Meta Muse has the top score of 9.3 among 12 scored products, based on 8/15 dimensions. In micro1 PersonalAgentBench, Gemini Spark completed 6 of 10 workflows with trusted completion (60%, 95% interval 31–83%); the intervals of all tested products overlap, so the ranking is not statistically separated. In SuperCLUE-XClaw's Aug 2026 batch of Chinese-language tasks, Wenxin Assistant (task mode) leads with 97.62/100; products within one point share a tier. PersonalAgentEval reports 60.5% resolved for Instinct across 152 scoring units from 132 task chains; this is a separate simulated-account evaluation.
Does PAEval rank personal AI agents?
Not yet. The PAEval Index is in development and no PAEval score or ranking is published. The site shows each independent evaluator's results on that evaluator's own scale, next to documented capabilities, and never averages scores across evaluators.
Where does the PAEval data come from?
From 344 cited sources: independent product-level evaluations, hands-on reviews, and vendors' official documentation, pricing and policy pages. PAEval tracks 45 products, 30 with full profiles, and 539 capability records. Each record keeps its source URL, date, evidence type and any region or plan conditions. Unknown is recorded as unknown, never as unsupported.
Can I download or cite the data?
Yes. Products, scores, capability records and sources are downloadable as JSON and CSV from https://paeval.com/about#data. Cite as: PAEval, snapshot 2026-10-28, https://paeval.com. When quoting a score, also cite the original evaluator.