Run your workshop on a company with an answer key.
For people who run workshops. Your attendees build agents against one simulated company - its Salesforce, Zendesk, Jira, Slack, Gong and the rest - and you grade what they hand in against answers Era computed when it built the company. Apply below for an allowance for the workshop.
A normal Era account has a small monthly call allowance. An approved workshop gets an allowance sized to the head count, for every attendee, for the length of the workshop.
Share the company with your attendees and read its request log: every call with the attendee who made it, the system, the tool or route, status and latency.
Era plants the facts and computes the answers before anyone asks. You get the key; attendees don't. A submission is right or wrong by comparison, not by a model's opinion.
* Unlimited means we size the allowance so your workshop doesn't hit it, not that there is no cap. Tell us the head count and the hours, and we set it with room to spare. If a run needs more mid-event, write to era@eon.io and we raise it.
From application to graded submissions.
- We approve and size it. We confirm the dates and head count, raise the allowance, and send you a sign-up link for attendees.
- You build the company. On Generate a company, pick the industry, size, model and systems. Everyone works against this one company, so everyone answers the same questions.
- You share it. From the company's card in Environments, share it with each attendee as a collaborator, up to 20 per environment. Collaborators issue their own tokens, so each attendee's calls carry their own name. For a bigger workshop we set up one environment per group, built the same way.
- We send you the key. You get the answer key for the company and day you hand out, privately. It never sits in the attendees' environment.
- Attendees build; you watch. Read Consumption while they work.
- You grade. Collect one answer per question and compare it with the key.
See how each attendee got their answer.
As the owner of the environment you read its request log on Consumption. It shows recent calls, newest first, and you can filter by system, status, method, tool and text.
- Which attendee made it, by the email you shared with
- The system and whether it came over the API or MCP
- The MCP tool, or the HTTP method, route and path
- Status and latency, so you can spot a stuck agent or a retry loop
- Request and response size, one click away
- Who reached the right answer by reading the right records, and who guessed
- How many calls an answer took, which you can use as a tiebreak
- Who is blocked: no calls, or a wall of 401s and 404s
- Who tried to write: environments are read-only, so a write comes back refused, as a 4xx in the log
Collaborators can also move the company to another day, and the answers change with the day. Grade against the key for the day you handed out, and check the day on the company card before you grade. Keep the shares until you have graded: once you revoke one, that attendee's past calls show an account id instead of their email.
The answers exist before the questions do.
Era doesn't label a company after the fact. It generates the business first, makes every system agree with it, plants specific records for questions to find, and then works out each answer from the records it wrote. Because the company is generated, not sampled from anyone's data, the key is computed over every record, not a labelled sample.
{
"id": "sf_open_pipeline",
"connector": "salesforce",
"kind": "value",
"question": "What is the total
open-opportunity pipeline amount?",
"value": 412500
}Each entry names the question, the system it lives in, what kind of answer it takes, and the answer. Cross-system questions join records from several systems.
{
"sf_open_pipeline": 412500,
"zd_open_urgent": 7,
"ghost_account_revenue": null
}One answer per question id. null means "I can't answer this from the data", which is
the right answer when the question asks about something the company doesn't have.
| kind | right when |
|---|---|
| exact | The count, amount, name or id matches the key. Numbers compare as numbers, with a tolerance where the question allows one. |
| set | The submitted items are exactly the key's items, in any order. Partial credit is the F1 of the two sets. |
| sequence | The items are the key's items in the key's order, such as a ranking. Partial credit is the share of positions that are right. |
| mapping | Every group maps to the right value, such as owner to closed-won amount. Partial credit is the share of groups that are right. |
| abstain | The attendee declined to answer a question the data can't support. Making up a number is wrong. |
Each graded answer gets a strict right or wrong and a score between 0 and 1, so you can rank by either.
Three workshop formats.
Teams wire an agent to Salesforce, Zendesk and Slack over MCP and answer 10 questions. The leaderboard is exact answers, then fewest calls.
How many open urgent tickets are there?
How many accounts have both open pipeline and an open, urgent support ticket?
The morning is single-system counts. The afternoon needs joins, rankings and medians across systems, then moving the company forward a day and answering again.
Which opportunity owner has the greatest total closed-won amount?
What is the median closed-won opportunity amount?
Mix in questions the data can't answer and documents with planted secrets and personal data. Score abstaining correctly and finding the planted findings, and read the log for refused writes.
Which synthetic secret findings are present in the stored documents?
The questions above are real questions from one of Era's companies. Claude Build Day attendees built on Era. To try a company before you apply, see the docs.