Help Expand PersonalAgentEval

We are expanding the PersonalAgentEval leaderboard, and we’d love your contributions! Has a personal agent ever let you down, missed an important detail, or failed to complete a task? Share your tasks and experiences to help us turn real-world failures into meaningful evaluation cases.

Our prior related work has received notable recognition, including an invitation to submit to Nature and acceptance to the NeurIPS 2026 ED track.

Every submission is reviewed and scored by a reviewer, with particular attention to the task description, the length and relevance of personal and environment context, and the number and detail of rubrics. Please contribute diverse domains across your submissions. Detailed review rules and authorship criteria will be published later.

Contribute to PersonalAgentEval

Share a failure

Did a personal agent get something wrong for you? Show us. Real failures become new test cases on the leaderboard.

4 steps · 5-10 min est per submissionNo accountRemove sensitive information before submission.

Basic info

Your contributor details, agent, and user profile.

Contributor name *

Your contributor name is only a display name on the contributor board. Your email is your unique contributor identifier and links your submissions. Use the same email and exactly the same contributor name for every submission. Different contributors may share a display name.

Email *

Your email uniquely identifies your contributions and is used for follow-up and authorship discussions. It is not displayed publicly. Always use the same email and contributor name together.

Which agent?

Personal agent

General agent

It could use (optional)

When did it happen (optional)

User profile *

Complete the profile fields below. MBTI is optional; “Prefer not to say” is available for the other fields.

MBTI (optional)

EnergyHow do you prefer to recharge?InformationHow do you prefer to process information?DecisionsHow do you tend to make decisions?LifestyleHow do you prefer to organize your life?

MBTI: not selected

Occupation *

Select your occupation. Choose Other to enter your own answer.

Occupation subdomain (optional)Gender * Age range * Main country of residence * Main languages used *

Select all that apply. Add other languages below, separated by commas.

Other languages (optional)

The task

What you asked for, and what about you it needed to know.

Domain *

Which domains best describe your case? Select all that apply.

Subdomain (optional)

User request / Task *

Personal and environment context — All *

Provide at least 5 relevant images OR at least 100 English words or 150 Chinese words in Paste text. This is a key focus of reviewer assessment. Share links may supplement your materials but do not replace this minimum. Chinese text is counted using word segmentation.

Images can include screenshots from any app, photos of the physical world, and wearable data such as activity, sleep, or sensor records. Provide relevant information that helps explain your personal and environmental context, and remove sensitive details before uploading.

You can combine images, a share link, and pasted text. Provide all the background the agent should know, including long context, earlier conversations, records, and preferences. Use images, an accessible share link, or pasted text. Identify the key personal facts in Personal and environment context — Core below and reference the supporting materials.

Paste a image (⌘V), drop it here, or click to choose Images, up to 10 MB each

In the chat, press Share, then Copy link.

Personal and environment context — Core *

What personal and environment context would any agent need to complete this task correctly and satisfy you? List essential preferences, constraints, intentions, or facts, regardless of the agent or its failure.

Where does this information come from in a real personal agent context? It may be explicitly stated in a record (e.g. an email or note), or inferred from behavior and other clues. Specify the source and, for an inference, the evidence and reasoning. Choose a suggestion or write your own.

What went wrong?

What happened, and what right looks like. This is what we will test other agents on.

What went wrong, and why do you think it failed? *

Select any relevant failure types, then describe what the agent did wrong and why you think it failed.

Other / custom failure tags (optional, comma separated) Failure description and reason *

Which context did the agent miss or misuse? (optional)

What should it have done? *

For example: stay under my $450 travel limit, and ask before booking.

Success criteria / rubrics

Add clear criteria and weight each with points (default: 10; no fixed total). Total earned score: below 10 = unacceptable, stop using personal agent; 10+ = acceptable, keep using personal agents. More points = greater user satisfaction.

We especially encourage detailed rubrics for implicit preferences (preferences not explicitly stated), proactive behavior (when and how far the agent should act on its own), and privacy and safety boundaries. These aspects will receive additional weight in reviewer assessment.

Example: The airfare must be under $450, or below the average airfare for the same route and travel dates.

Available points: 10 · Minimum acceptable earned score: 10

What makes this challenging? (optional)

We especially welcome cases requiring initiative, memory, or several sources. Everyday failures are welcome too.

Other / custom challenge tags (optional, comma separated)Why is this case challenging? (optional)

Could someone who doesn't know you get this right? (optional)

This tells us whether the task really needs a personal agent.

Review your submission

Review your contributor details, profile, task, all context materials, core facts, failure explanation, and weighted rubrics below. Use the edit buttons to return to a step. Submitting sends the reviewed text, images, profile, and contact details to the contribution server. A submission ID is shown only after the server confirms saving.

Submission overview

Review materials and sensitive information

Detection supports multiple languages, including Chinese, but may miss sensitive information. Review all materials yourself; shared-link contents are not scanned.

Hidden

Additional sensitive information to redact and remove (comma separated)

Separate names, words, or phrases with commas. Exact matches in any language (including 中文) are hidden in text and recognized image text. Draw boxes on images to hide anything missed; check shared links separately.

✓

Submission received

Your submission has been saved for review. Keep the submission ID below for follow-up.

Contributor board

About the contributor board

Accepted cases are submissions approved after review. Acceptance rate is accepted cases divided by total submissions. Rankings use accepted case count, then acceptance rate.

What we received

Submissions are stored on the contribution server for review. Your email identifies your contributions and is not shown publicly. Automatic redaction is incomplete; check all materials before submitting.