Arhaan Aggarwal
CEO & co-founderImplemented computer-use evaluators and verified end-to-end workflows through OSWorld-v2.
LinkedInSimulations to test and train AI agents.
Realistic environments built around your tools, data, permissions, and workflows. Today, we help teams test what their agents actually change.
A reply isn’t
a result.
A refund request crosses the support desk, order system, and payment ledger. A timeout can turn one payment into two.
In development: environments and training methods that help agents learn from these failures.
Customer request“Refund my $120 order.”Payment recorded. Response timed out.
Unsafe retry$240.002 refunds recorded
A new request key pays again.
Guarded retry$120.001 refund recorded
The same key preserves one payment.
Synthetic data. Scripted paths. No model called.
See what changes.
The task: refund the right customer, once, within the rules. Choose a failure condition and see what the agent changes.
Interactive example · synthetic dataScripted paths. No live model or production calls. Not a customer run.
What happened
0 EVENTS- Ready when you are.Run the scenario to follow the actions and inspect the records.
A check passes only against this example’s defined rules—not as a general safety guarantee.Three failure conditions. Two paths each.
Your workflow.
Its whole world.
Custom environments, built and delivered directly with your team. Accounting and ERP, GTM, or customer support: we recreate the detail that matters to the task.
Scope an environmentDefine “right.”
Map the task, connected tools, permissions, and outcomes your agent needs to get right.
Recreate the systems.
Model the data and dependencies at the level of detail the workflow needs.
Make failure observable.
Introduce timeouts, stale records, customer boundaries, and approval requirements.
Check the actual result.
Independently verify task completion, unintended changes, and whether the evaluation itself can be trusted.
From evaluation
to learning.
Customer work reveals consequential failures. With appropriate permission, those patterns can inform verified task families, training environments, and adaptive curricula.
- Available today
Custom environments
Customer-specific workflows for agent evaluation, scoped and delivered with your team.
- Next
Maintained environments
Refresh environments and evaluations as agents, tools, and workflows change, through recurring engagements and reusable components.
- Planned training offering
RL environments
Verified task sets and reinforcement-learning environments where developers and model labs can train agents through repeated practice.
- Research direction
Adaptive curricula
Self-play-inspired training that uses verified failures to propose the next challenge. An applied AI lab developing environments and methods for unfamiliar workflows.
Planned training loop
Challenges evolve. Independent checks stay central.
Challenge
One AI proposes a workflow task.
Attempt
Another agent attempts it in the environment.
Verify
Independent checks validate suitability, actions, outcomes, and rewards.
Adapt
Verified failures guide subsequent challenges.
Our first research question Can adaptive curricula improve performance on held-out workflows compared with fixed curricula, at comparable total cost?
This is planned work, not a demonstrated training or generalization result. Customer data and task reuse require appropriate permission.
Built and delivered.
Three custom environments.
One pilot customer.
Delivered for testing internal products and tools. Evidence of delivery, not measured performance or business impact.
Meet the founders.
Experience across agent evaluation, synthetic data, and B2B customer discovery.
What needs
to go right?
Tell us about the workflow, the systems it touches, the task to complete, and the failures that matter.
Discuss your workflow Contact Arhaan on LinkedIn · opens a new tab