PAEval
Data as of 9 Oct 2026
Capabilities / Reliability & recovery

Personal AI agent reliability and recovery

As of 9 Oct 2026, 17 of the 30 profiled personal AI agents have documented or observed records for retries, recovery, monitoring and failure handling. 1 of them have independent observations; the rest rest on vendor documentation or partial evidence. 10 profiled products have no record yet, which means unknown, not unsupported.

Availability is not task success: a documented capability says nothing about how reliably the agent completes the task. For measured performance see the leaderboards.

Independently observed (1)

Shuffle

DNFT · Creation/entertainment companion assistant, including website execution

  • Third party alternative routes Independently observed
    On 9/9, I switched to Booking.com after booking an error; Amazon failed to switch to a brand store but got stuck in the identity check; I couldn't claim that the purchase was completed.source ↗

Documented by the vendor (2)

OpenAI dots

OpenAI · Persistent personal agent

  • Activity and handoff Documented
    Activity displays ongoing/completed tasks, files and pending tasks; supports taking over and changing priorities; task completion does not automatically prove that the results have been delivered.source ↗
  • Stop scope Limited or conditional
    Pause only stops the main task; assigned tasks and future plans must be stopped separately, and external actions taken will not be rolled back.source ↗

Town

Town.com · Work-based personal agents and team automation

  • Breakpoint recovery Documented
    The routine caches results step by step and reuses completed steps after approval or temporary errors; caching is limited to this run.source ↗

Partly disclosed or tested (5)

Instinct

Instinct (Spear Street Technology) · Universal Personal Execution Agent

  • Third-party end-to-end coverage Partly disclosed / partly tested
    Travel, shopping, and email on the list have action results; chain tasks are not scored, and the reliability of complex tasks cannot be derived from a single performance.source ↗

szn

Nomadic Futures · Universal Life Execution Agent

  • Task record Partly disclosed / partly tested
    The policy saves task status, results, logs and action records; recovery/idempotence/failure retry details are not disclosed.source ↗

Tomo

Mapo Labs · Personal goals/life companionship and execution assistant

  • Third-party multi-step gap Partly disclosed / partly tested
    The 9/10 test was able to staging air tickets and asked first because it was over budget; the email/Drive was not connected, and check-in could only recommend manually providing information.source ↗

Caddy

Caddy Technologies Inc. · Message-Based Universal Personal/Home Execution Agent

  • Confirm completion and unknown mechanism Partly disclosed / partly tested
    The travel page uses final confirmation from the booking platform as the booking criteria; browser login/store availability will limit execution. Idempotence and breakpoint recovery are not disclosed.source ↗

Cue by Manus

Manus · Independent identity, multi-agent collaborative personal execution agent

  • Agent relay Partly disclosed / partly tested
    Official research site → candidate list → handover of recommended deck; error recovery, idempotence, long task timeout and approval reentrancy specifications are not disclosed.source ↗

Limited or conditional (9)

Meta Muse

Meta · Persistent personal agent

  • Observability and recovery Limited or conditional
    The browser can be viewed and taken over, with action auditing; the VM is continuously backed up and can be restored, officials say.source ↗

Grok Bot

xAI (via Cursor) · Individual/Team Agent

  • Recovery Limited or conditional
    Observe the computer, retry/Recover/Update; restore to retain persistent files/login, Reset to the latest snapshot may lose recent changes, and temporary software packages are not guaranteed to be durable.source ↗
  • Enterprise observability Limited or conditional
    Consumption session display tool activities; enterprise Audit logs, Action Recording, and OpenTelemetry are set separately and depend on the package, and Action Recording is turned off by default.source ↗

Claude (including Cowork execution)

Anthropic · General-purpose work agent

  • Audit and recovery Limited or conditional
    Cloud session isolation; Team/Enterprise can use OpenTelemetry, and Compliance API can capture Cowork activities; file/web page tools still run when the local VM is unavailable, and code execution reports workspace…source ↗
  • Disclose evidence of failure Limited or conditional
    2026-09-08 User reported that Windows local scheduled task permissions have returned, providing version, log and reproduction steps; the issue is still Open when opened. It is only a first-hand failure report of a…source ↗

Manus 2.0 / Manus Studio

Manus · General-purpose work agent

  • Recovery and Observability Limited or conditional
    The cloud browser can watch in real time and continue after manual takeover during verification code/multi-factor authentication; Cloud Computer has Web Terminal/SSH, CPU/RAM/storage monitoring; process logs can be…source ↗
  • Independent hands-on observation Limited or conditional
    The SEW author used Manus 2.0 to create a 30-second commercial on September 30: the first draft took 21 minutes, and the subsequent three revisions took less than 4 minutes and were called completed.source ↗
  • Manufacturer efficiency claims Limited or conditional
    Compared with the old system, Cascade claims to reduce tokens by 23.2%, time consumption by 28.2%, and operating costs by 32% under a test configuration.source ↗

Lindy(Personal Lindy + Teammate)

Lindy · Individual/Team Agent

  • Recovery/Testing/Observability Limited or conditional
    Routines have a dry run that does not generate external writes; personal routines can be converted into Python scripts, and Lindy can be awakened for repair/testing if the script fails. Connectors display…source ↗
  • External data quality Limited or conditional
    Hlido gave 73/100 and 4/5 public surface inspection passes, but the self-reported non-login test behavior uses automatic web page inspection and AI writing; it cannot be used as a personal agent task completion rate.…source ↗

Gemini Spark

Google · Persistent personal agent

  • Observable and recoverable Limited or conditional
    The task work panel displays steps, plans/progress, and read and write files; it can be stopped or taken over. Temporary Mac backups are only kept until new tasks or 24 hours, and cannot be considered a full version…source ↗

Perplexity Computer (Cloud)

Perplexity · Cloud execution agent

  • Observe/Recover Limited or conditional
    There is task progress, usage of each thread and automation history; normal operation can be suspended due to insufficient points and resumed after replenishment; automation will be suspended if it fails…source ↗

Perplexity Personal Computer (local access)

Perplexity · Cloud + local execution agent

  • Stuck/recovery Limited or conditional
    The official recommendation is to first go to the web to determine whether only the desktop UI is stuck, and then restart the app; if it is really stuck, you can stop it, check permissions, check that modifications…source ↗

Microsoft Copilot Cowork for personal accounts(Preview)

Microsoft · Personal Execution Agent (Preview)

  • Observe/Interrupt/Resume Limited or conditional
    Skill activities/progress are displayed step by step, and directions can be added during execution, paused, resumed/cancelled immediately after the next step, and files can be reviewed in the output panel. No full…source ↗

Planned (1)

Pally

Pally Technologies · Multi-channel universal personal execution agent

  • Restore status Planned
    The update page lists shopping cart recovery, call silent line handling, etc. as Actively working on, which should not be translated as verified stable capabilities.source ↗

Not disclosed (2)

Ollie

Confabulation Corporation · Home personal assistant

  • Official multi-step recovery/retry mechanism Not disclosed
    The listed sources were reviewed, but sufficient public documentation was not found.source ↗

Asaply

Asaply Inc. · Local consumption/errand transaction type personal agent

  • Failure recovery and exception chargeback execution specifications Not disclosed
    The listed sources were reviewed, but sufficient public documentation was not found.source ↗

No record yet

OpenClaw, Hermes Agent, AutoGLM (hosted application), Doubao Phone Assistant (consumer version), Qwen App (task execution / work assistant), Coze 3.0 / Coze Agent (formerly Coze Space), HONOR YOYO Agent, MiniMax Agent/MiniMax Code Work Mode, QwenWork, WorkBuddy (Tencent)

Other capabilities

Personal AI agents that work in the background Personal AI agents with cloud or local computers Personal AI agents that make calls and send messages Personal AI agents with app connectors and tools Personal AI agents with long-term memory Personal AI agents in chat apps, voice and other channels Personal AI agents for research and deliverables Personal AI agents that book and buy Personal AI agents for teams and parallel work Personal AI agent permissions, privacy and security