According to Eon's paper, Era is built around a complete fictional company, with product simulators, internal databases, benchmark questions, and computed answer keys all tied to the same setup. The company says a single seeded entity graph supplies shared data to simulators of Salesforce, Zendesk, Slack, Gong, and other products, so the different systems describe one consistent business instead of a pile of disconnected mocks.
That matters because one of the harder parts of testing enterprise agents is not just asking whether a model can answer a question, but whether it can do so while navigating the same kind of cross-tool sprawl that real employees deal with. In a post describing the product, developer Santiago Valdarrama said Era can generate a synthetic company across multiple systems including CRM, tickets, Slack, files, emails, and call recordings, then let developers reset the company to its initial state and rerun tasks. That description is from a reaction post, not the paper itself, but it lines up with Eon's broader pitch for repeatable agent testing.
How Era says grading works
A big part of the claim here is not just the simulated company, but the scoring. Eon's paper says every expected answer is computed from the final records, so grading is exact. In other words, the benchmark is supposed to know the right answer because the system generated the underlying company data in the first place.
That avoids a familiar benchmark problem: using another model, or a human reviewer, to decide whether an agent's answer was good enough. A separate reaction post summarized the same point by saying Era wrote every record, so code can compute each question's exact answer without relying on a human or model judge. Again, that's commentary on Eon's work, not independent verification, but it matches the mechanism described in the paper.
What the paper reports
Eon's paper reports a couple of headline numbers meant to show the synthetic company setup looks more realistic than a rough first pass. Across 23 generated companies, the authors say the mean realism score increased from 61.8 to 97.0, with zero records flagged as synthetic by their adversarial detector.
The benchmark results are more mixed, which is probably the interesting part if you're actually trying to compare models. In the paper's simulator-track comparison, nine models answered the same 33 questions three times each. Reported accuracy estimates ranged from 42.4% to 76.8%. The paper also says three of 36 pairwise differences remained supported after correction, which suggests some apparent gaps between models did not hold up once the authors adjusted for multiple comparisons.
That doesn't make the benchmark weak, but it does make the results less cartoonishly simple than a leaderboard screenshot. The setup may be useful precisely because it produces a narrower, more debatable spread than cherry-picked demos do.
Why people building agents are paying attention
The reaction from people in the AI and enterprise software orbit has been broadly positive. Omar Shams said on X that simulated companies are a big deal for improving agentic products, framing Era as useful because good agent evals need environments that behave like real companies. Box CEO Aaron Levie made a broader version of the same argument, writing that a lot of agent deployment is bottlenecked by the ability to test, tune, and optimize agents in realistic work environments.
That reaction helps explain why a project like this is getting attention. It does not, by itself, prove Era's product claims or benchmark quality. What the supplied evidence supports is narrower: Eon has described a simulated enterprise benchmark with shared data across common business tools, exact-answer grading, and resettable test conditions, and its paper reports specific realism and accuracy numbers. Whether that turns into a standard way to evaluate enterprise agents is still an open question.
