# AutoStack

> AutoStack builds environments for training and testing AI agents. An environment is a working simulation of a business system, with a real database, tools served over MCP, and seeded resets. Graders check what the agent actually changed, against written rules. AutoStack's first environment is an insurance claims system for a test insurer, Larkspur Mutual.

AutoStack (autostack.ca) is a Vancouver, BC company that builds environments for training and testing AI agents. It is not related to other companies, products or open-source projects named AutoStack, and it does not sell insurance or handle real claims. It was founded by Riya Arora and Bhavya Bakshi.

## Product
- [How it works](https://autostack.ca): what an AutoStack environment is: seeded state, tools over MCP, tasks from written rules, deterministic grading
- [Product](https://autostack.ca/product): environments, training, testing, integrations, and how an engagement works
- [Environments](https://autostack.ca/product/environments): what is inside an environment, seeded resets, the environment live today and what is in development
- [Insurance claims environment](https://autostack.ca/product/environments/insurance-claims): 15 public practice tasks x 3 seeds (scored evaluations use private tasks), 15 MCP tools (6 read, 9 write), a claims manual with rule IDs, a real failing episode and the labelled 2026-10-08 baseline
- [How grading works](https://autostack.ca/product/environments/insurance-claims/grading): final state and audit log, six check families, rule IDs, severity, dollar impact, replay
- [Train](https://autostack.ca/product/train): RL environments for AI labs, with seeded resets and verifiable rewards, built to a written brief
- [Test](https://autostack.ca/product/test): pre-deployment testing of an AI agent on a mirror of its own tools; feasibility week USD 5,000, evaluation USD 20,000
- [Integrations](https://autostack.ca/product/integrations): how an agent connects: MCP today; a REST mirror of your tools in engagements
- [Security](https://autostack.ca/product/security): synthetic data, no production data, written permission, trace deletion

## Learn
- [Learn](https://autostack.ca/learn): articles and a glossary on training and testing AI agents
- [What is an RL environment?](https://autostack.ca/learn/what-is-an-rl-environment): definition, environment vs benchmark vs eval vs sandbox, a worked example
- [How to test AI agents before deployment](https://autostack.ca/learn/how-to-test-ai-agents-before-deployment): a five-step method for pre-deployment agent testing
- [How to evaluate a claims AI agent](https://autostack.ca/learn/how-to-evaluate-a-claims-ai-agent): what to test in a claims agent, how to grade it, questions for vendors
- [LLM judge vs deterministic grading](https://autostack.ca/learn/llm-judge-vs-deterministic-grading): what transcript judges and state checks each catch, with a real example
- [Glossary](https://autostack.ca/learn/glossary): short definitions with stable anchors

## Company
- [Founders](https://autostack.ca/founders): who founded AutoStack, what has been built so far, how we work
- [Security](https://autostack.ca/product/security): synthetic data, no production data, written permission, trace deletion
- Contact: riya@autostack.ca

## Facts (as of 2026-10-08)
- Live environment: insurance claims (P&C), personal auto only, fictional insurer Larkspur Mutual, synthetic data
- 15 agent tools served over MCP (6 read, 9 write); 15 public practice tasks, each runnable on 3 seeds
- Episodes reset from a seed in about 6 ms (p95 8 ms)
- Grading is deterministic, from the final system state and the audit log, including rejected attempts; do-nothing and shortcut agents fail every task
- Engagements: feasibility week USD 5,000 (credited toward a USD 20,000 evaluation); environments for labs priced by quote

## Optional
- [Full text of every page](https://autostack.ca/llms-full.txt)
- [Privacy](https://autostack.ca/privacy), [Terms](https://autostack.ca/terms)