{"slug": "computer-work-val-v1-0", "title": "Computer Work Val v1.0", "summary": "Computer Work Val v1.0, a benchmark of 300 tasks drawn from 10 domains and 36 roles, was released to test and train AI agents on simulated digital work, with each task built from high-fidelity clones of tools such as Jira, Google Sheets and Salesforce seeded with mock data from the person who performs the job. The environments run on 210 simulated apps, including Gmail, Excel, Slack, Salesforce, Jira, Zendesk, Stripe and Figma, alongside open-source systems such as ERPNext, OpenEMR and Odoo. Contributors were drawn from US occupations with the largest share of labor costs in digital work and passed a five-stage pipeline — initial screen, interviews, task creation training, a three-hour work trial and final human review — with tasks checked on four quality dimensions including realism.", "body_md": "The contributors come from the US occupations with the largest share of labor costs in digital work. Each one turns a piece of their own job into an environment. **Computer Work Val v1.0** is 300 of those tasks, chosen to represent every domain, and the environments can be used for both testing and training.\n\n## Results\n\n## How the environments are created\n\nWe build each environment from high-fidelity clones of tools like Jira, Google Sheets and Salesforce, seeded with mock data from the person who actually does the task.\n\n### Simulators used\n\nThe environments run on 210 simulated apps, from Gmail and Excel to open-source ERPs, health records and helpdesks.\n\n- Gmail\n- Google Sheets\n- Slack\n- Word\n- Excel\n- PowerPoint\n- Salesforce\n- Google Drive\n- Jira\n- DocuSign\n- HubSpot\n- Shopify\n- Google Docs\n- Canva\n- Zendesk\n- Stripe\n- Figma\n- Power BI\n- Trello\n- Google Meet\n- Freshdesk\n- Amazon\n- BambooHR\n- Fathom\n- Airtable\n- 7shifts\n- TurboTax\n- DoorDash\n- Airbnb\n- Freshdesk (MCP)\n- Xero\n- Booking.com\n- Jira Service Management\n- Zoho Inventory\n- Notion\n- Splunk\n- Zoho CRM\n- PagerDuty\n- UpKeep\n- craigslist\n\n## Show all 210\n\n- Grist\n- BookStack\n- Mattermost\n- Docspell\n- Nextcloud\n- Metabase\n- ERPNext\n- WordPress\n- Baserow\n- WooCommerce\n- OpenEMR\n- Plausible\n- Odoo eLearning\n- ONLYOFFICE\n- Znuny\n- Bahmni\n- Worklenz\n- GLPI\n- Superset\n- Akaunting\n- Bugsink\n- osTicket\n- Part-DB\n- GitLab\n- Plane\n- listmonk\n- Probo\n- Odoo\n- Odoo Manufacturing\n- Grafana\n- EspoCRM\n- Forgejo\n- Formbricks\n- YAFFA\n- Redmine\n- CloudBeaver\n- OpenCart\n- eLabFTW\n- Odoo Maintenance\n- Datashare\n- Odoo Invoicing\n- Horilla\n- Kanboard\n- NetBox\n- authentik\n- CKAN\n- Snipe-IT\n- Dittofeed\n- Frappe HR\n- Dolibarr\n- Docs\n- Odoo Project\n- CourtListener\n- Memos\n- GlitchTip\n- Uptime Kuma\n- ownCloud\n- Odoo Purchase\n- Healthchecks\n- Odoo Inventory\n- Documize\n- Documenso\n- Bigcapital\n- Keycloak\n- DocuSeal\n- Easy!Appointments\n- FreeCAD desktop\n- Moodle\n- Airflow\n- Airbyte\n- OrangeHRM\n- GRR\n- Answer\n- Cascadia PLM\n- FreeScout\n- Bitcart\n- Nextcloud Office\n- doccano\n- Redash\n- Firefly III\n- Mifos\n- EveryDocs\n- openIMIS\n- Lychee\n- Axelor\n- Ghostfolio\n- Odoo Expenses\n- LocalStack\n- bewCloud\n- tillywork\n- Invoice Ninja\n- Odoo Employees\n- CARE\n- b1gMail\n- Odoo Sales\n- Econ market\n- Bytebase\n- docassemble\n- GitHub\n- BookLore\n- Tumble\n- n8n\n- Kimai\n- Chatwoot\n- Wiki.js\n- NocoBase\n- InvoiceShelf\n- VICIdial\n- Odoo Email Marketing\n- DAViCal\n- Hyperswitch\n- iDempiere\n- Leantime\n- Mindustry\n- Twenty\n- Zulip\n- Bots EDI\n- Hi.Events\n- Kitchen 3D\n- IceHrm\n- alf.io\n- InvoicePlane\n- Penpot\n- Docmost\n- Flare\n- Crater\n- Bytepack\n- Floor 87\n- Lemmy\n- Seafile\n- Parallax\n- Automatisch\n- Vexa\n- Odoo CRM\n- Frappe Insights\n- OpenAlgo\n- KoboToolbox\n- Gridfall\n- Jeedom\n- OLDP\n- 2FAuth\n- ArchivesSpace\n- Cabot\n- Halo\n- DreamFactory\n- Odoo Events\n- Accent\n- µTask\n- TYPO3\n- Blender\n- Kestra\n- AFFiNE\n- 4ga Boards\n- vutuv\n- Odoo Time Off\n- Ever Gauzy\n- CodiMD\n- Aleph\n- Ballerine\n- Databunker\n- Peppermint\n- SuiteCRM\n- Cachet\n- Etherpad\n- Admidio\n- ChiefOnboarding\n- Aptabase\n- Drupal\n- Krayin\n- Coder\n\n## How we select people\n\nEvery task contributor goes through a rigorous pipeline to ensure they meet our quality bar.\n\n1. 1#### Initial screenWorks at a relevant organization · uses AI heavily in their job · 3+ years in their domain\n2. 2#### InterviewsA communication skills assessment and a domain expertise assessment\n3. 3#### Task creation trainingAn introduction to RL task creation and how to create high-quality tasks, then Plato platform training\n4. 4#### Work trialA three-hour work trial\n5. 5#### Final human reviewDomain experts and a Plato RL task specialist check accuracy, relevance, task quality and fit with RL objectives\n\n### Who the tasks come from\n\nThe 300 tasks span 10 domains and 36 roles.\n\n## How we ensure world quality\n\nWe check every task on four dimensions to make sure it reflects real work, poses a meaningful challenge and is worth getting right.\n\n#### Realistic\n\nThe environment and prompt match the author’s real job.\n\n- The software they actually use\n- Anonymized data, in the places it would normally sit\n- A short prompt, the way work actually arrives\n\n**In the Juniper Bean example** The accountant’s own stack: Excel, Gmail, Stripe and Mercury. Last quarter’s P&L sits in an email thread, where it normally would.\n\n#### Fair\n\nEverything needed to finish the task is in the environment.\n\n- Information is where someone in the role would look\n- Conflicting data has a clear source of truth\n- Each grading criterion is objective and tied to what the task requires\n\n**In the Juniper Bean example** Stripe shows every CAD payment next to its USD-converted amount, so the right revenue figure is findable.\n\n#### Difficult\n\nThe work takes reasoning and several steps, not a lookup.\n\n- Reconciling information across sources\n- Working through conflicting records\n- Difficulty from the work itself, never from missing information\n\n**In the Juniper Bean example** Revenue spans USD and CAD. Summing the Stripe export at face value overstates it by $3,180.\n\n#### Severe\n\nA wrong answer has real consequences.\n\n- Financial or legal impact, or harm outside the team\n- Hard to reverse or likely to go unnoticed\n- Actions taken without the required approval\n\n**In the Juniper Bean example** The P&L goes straight to Elena, the CFO. An inflated revenue line reaches leadership unnoticed.\n\n**Reviews:** two experts from the same occupation agree on each task’s realism and fairness, and the Plato team reviews it for environment defects, reward hacking and prompt/grading mismatches.\n\n## Want access?\n\n[Partner with us](https://plato.so/partner/) to get access to any of these environments.\n\n## Work with us\n\nWe’re hiring! Check out [open roles](https://jobs.ashbyhq.com/plato.so) or email [hi@plato.so](mailto:hi@plato.so).", "url": "https://wpnews.pro/news/computer-work-val-v1-0", "canonical_source": "https://plato.so/research/computer-work-val/", "published_at": "2026-10-28 04:49:12+00:00", "updated_at": "2026-10-28 05:12:25.360079+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-agents", "ai-research", "ai-tools"], "entities": ["Computer Work Val v1.0", "Jira", "Google Sheets", "Salesforce", "Gmail", "Excel", "Slack", "Plato"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/computer-work-val-v1-0", "markdown": "https://wpnews.pro/news/computer-work-val-v1-0.md", "text": "https://wpnews.pro/news/computer-work-val-v1-0.txt", "jsonld": "https://wpnews.pro/news/computer-work-val-v1-0.jsonld"}}