# How Eon built Era with Devin: simulated companies for training and evaluating AI agents

![Vimeo](https://devin.ai/_next/image?url=%2F_next%2Fstatic%2Fimmutable%2Fmedia%2Fcover.3-u6bnk0-98tq.webp\&w=3840\&q=75)

98%

Of merged Era PRs written by Devin

\>300

Devin sessions running for over a day

2,720

Devin sessions launched by other Devin sessions

## About the company

[Eon](https://www.eon.io/) is building the Autonomous Data Foundation for the AI era. Founded by the team behind AWS Disaster Recovery and Migration services, Eon has raised $500 million from leading investors, including Sequoia, Lightspeed, Greenoaks, BOND, and Elad Gil, and serves many of the world's leading brands.

Industry: Data Infrastructure**[Visit site](https://www.eon.io/)**

Eon backs up cloud data for enterprises: exabytes across hundreds of thousands of protected databases, buckets and disks. Eon also prepares that data for AI agents and analytics. Its managed platform combines its own storage with fast ingestion, which lowers costs and cuts the infrastructure customers have to run.

## What Era is

Operating at that scale taught Eon’s engineers how enterprise data actually behaves. Distributions are skewed, fields are half empty, five systems disagree about the same customer, and APIs page at 100 rows. Most synthetic data looks nothing like this, which is a problem for teams building agents and training models. Their agents have to work on messy data, but real enterprise data can’t be shared, so they train and evaluate on clean, single-system datasets and find the gaps in production.

Eon built Era to give those teams realistic data to work with. It generates a whole company, with customers, employees, deals, support tickets, chat messages and recorded calls, and serves it through 74 simulated systems, including cloud databases, buckets, file systems, and SaaS tools like Salesforce, Slack, Zendesk, Jira, and Gong. The same customer shows up consistently across all of them. Each system has the vendor's own schema, query language, auth, pagination and errors, so official SDKs and MCP clients connect unchanged.

Era contains no customer data: every record is generated. The generator is tuned to aggregate statistics about real enterprise data, like how skewed distributions are, how often fields are left blank and how data is spread across systems, rather than to any actual records or values.

Because the generator picks each answer first and then plants the records that produce it, every question has one exact answer that code can grade without a second AI model. To answer a question, an agent usually has to piece together facts from several systems. For example, when asked what service credit a customer is owed, the agent finds zero in Salesforce, a discount the account manager promised on a recorded sales call, and support tickets that show how much of that discount applies. Era can generate a new company for any industry and size, and each company's data keeps changing over time, so agents can't memorize the answers and teams don't run out of training data.

## How Eon’s team built Era with Devin

A small Eon team started building Era with Devin. Most of the work was making the simulators convincing enough that an agent can't tell them from the real product. Each of the 74 simulators has to match the real product’s data formats, query language, logins, page limits and error messages. Developers’ existing code and the vendors’ official toolkits have to connect without a single change. All 74 also have to tell the same story about one company, so a customer’s deal in Salesforce has to match their tickets in Zendesk and their calls in Gong.

Eon's team designed the architecture: every simulator serves the same simulated company, answers are decided before the data that proves them, and a fidelity loop tests each simulator against the vendor's own SDKs. Devin built the 74 simulators inside that design. Each pass through the fidelity loop generates data, compares it with the real tool, finds where an agent could tell it's simulated, and fixes it. Each round means waiting hours for tests and data builds to finish.

## Sessions that wait hours for tests

A coding agent running in a container loses its work when the container shuts down, or keeps paying for compute the whole time it waits. Devin saves the whole machine, including memory, running processes and files. The session shuts down while idle and wakes exactly where it left off when tests finish. 538 of Eon's sessions stayed open for more than 12 hours, and 327 for more than a day.

> "When I get on a plane, I tell Devin to keep auditing and fixing for the next few hours, and I land to merged PRs."
>
> — Benjamin Gruenbaum, Principal Engineer, Eon

## Thousands of sessions in parallel

Every Devin session also runs on its own virtual machine, and Devin can run thousands at once. A Devin session can split a big job, hand the pieces to other Devin sessions and merge what comes back. Devin ran 2,900 sessions on Era, and 86 of them launched the other 2,720. One audit of the simulators the team designed fanned out to 706 sessions, one for each simulator and type of record, and produced 123 merged PRs. On busy days, the team launched 99 sessions in an hour and 254 in a day.

Each session works inside its own PR. It waits on tests, fixes failures and pushes again until the PR is ready to merge. Devin wrote 2,600 of the 2,652 PRs merged into Era. Those PRs added about 33,800 commits, 3 million lines of code and 61,000 tests.

## Era's launch

Eon launched Era's public console on October 5. In its first five weeks it got hundreds of access requests, and outside users made about half a million calls. About 280 outside teams now run agents against it.

> "This would not have been possible without Devin - it really let us take our expertise and build on it quickly."
>
> — Benjamin Gruenbaum, Principal Engineer, Eon

## Related customer stories

- [How Modal uses Devin to investigate 80% of its incidents before an engineer opens the thread](https://devin.ai/customers/modal)
- [How Devin Oncall became Reducto's first responder, so engineers rarely get paged](https://devin.ai/customers/reducto)
- [How Crosby is building an AI-native law firm with Devin](https://devin.ai/customers/crosby)