P1
Reading screens
OCR and layout models read every frame down to the word, with a vision model for the rest. Dense tables and nested forms have to be read exactly.
Open questionRecovering table structure through merged cells and frozen panes.
Every business change ends up as changes to your systems, and each system has its own specialists. Mando’s agents know every application, learn how yours is set up, and make the change across all of them.
§1Thesis
Enterprise admins translate business changes into system changes. Today that takes a specialist for every application, each with their own queue and their own way of working.
Thesis. Agents can do this across any enterprise system, with built-in knowledge of each application and the ability to learn each customer’s environment.
(a)Today. A specialist per system, each with a separate backlog.
(b)With Mando. One set of agents across every system, sharing what it learns.
§2How it works
01 Learn
Agents read calls, tickets and your admins’ past work, then inspect each system’s screens and backend to see how it is set up today.
02 Build
Each agent works in its own browser with code execution. It plans the change and makes it on every screen it touches.
03 Prove
Nothing ships until it passes in a sandbox. The agent records the state before and after, and keeps what it learned for the next change.
§3Research
Screens are dense, state is hidden, and every action lands in a system people are working in. There is no reset button. These are the problems we work on.
P1
OCR and layout models read every frame down to the word, with a vision model for the rest. Dense tables and nested forms have to be read exactly.
Open questionRecovering table structure through merged cells and frozen panes.
P2
Personal and proprietary values are masked on every frame at ingest, before anything is stored or seen by a model.
Open questionCatching company identity hidden in configuration names and free text.
P3
Enterprise systems allow some changes only from some screens. We rebuild state from pixels and backend queries, and hold it across thousands of actions.
Open questionDetecting drift between what the agent saw and what actually committed.
P4
Each team’s history becomes a private record of how its systems were configured and why. The agent reads it before acting and adds to it after.
Open questionInferring why a rule exists when nobody wrote the reason down.
P5
Agents learn by reinforcement inside each customer’s environment, from every change they build and test.
Open questionCarrying what works in one module to a module the agent has never seen.
P6
One remote browser session, hours long. The agent gathers context through clicks and code, builds the change, then retests it.
Open questionClean rollback after a half-applied change, hours into a task.
P7
The unit of work is a verified change: the intent, the state before and after, and the runs that validate it.
Open questionScoring a change that is correct but was built a different way.
P8
If there is a problem in this space you would rather work on, tell us what it is.
§4Results
Our agents lead on SCUBA and WorkArena. Public benchmarks cover only a handful of systems, so we build our own for every system our agents run.
| Benchmark | From | What it measures | Tasks | Source |
|---|---|---|---|---|
| SCUBA | Salesforce AI Research | Admin, sales and service work in live Salesforce sandboxes, scored by milestone | 300 | arXiv 2509.26506 |
| WorkArena++ | ServiceNow Research | Compositional workflows on a live ServiceNow instance: planning, retrieval and memory | 682 | arXiv 2403.07718 |
| CRMArena-Pro | Salesforce AI Research | Sales, service and quoting in a Salesforce org, single and multi turn | 19 | arXiv 2505.18878 |
| EntWorld | Tsinghua University | ERP, CRM and IT service tasks, each with a verifiable end state | 1,756 | arXiv 2601.17722 |
| Ledger | NeurIPS 2026 | 53 modules of a production ERP, scored on what reached the database | 424 | neurips.cc |
| EnterpriseBench | Fujitsu Research | Engineering, HR, finance and admin tasks behind access controls | 500 | arXiv 2510.27287 |
| SOR-Bench Ours | Mando | Verified questions across 31 modules, written with practitioners. The first of one benchmark per system we run. | 2,214 | sorbench.com |
§5Careers
A small team of researchers and engineers working on real enterprise systems, with customers who need the change made correctly the first time.
Don’t see your role? Write to talent@mando.work with what you would want to work on.