Teach agents on a real-looking company

* Not literally unlimited: every account with a university or college address gets the allowance above, per person, instructors and students alike. A busy agent makes 30-50 calls a session, so a semester of coursework stays far below it. If your class gets close, write to us and we raise it. Calls are counted the same way as on every Era account - see usage limits.

Open a class and hand out its code. A student who joins gives you read-only access to every environment they own, for as long as they stay on the class:

• each environment's data, systems and which day it serves - the same view the student has, without the buttons that change anything
• calls this month and in total, per environment and per system
• the request log: every call their agent made, with its status and timing

You can't write to a student's company, mint its tokens, share it or delete it. The student is told what you can see before they join. Leaving the class, being removed from it, or the class closing ends your access at once.

$ era class create "Agents 101"        # prints the join code
$ era class show K7QM-3F9P             # roster, environments, calls
$ era class join K7QM-3F9P             # the student's side

Every student's company is different, so a fixed answer key doesn't work. In Era the ground truth is whatever the company's own systems say. An assignment names, for each question, the read-only call that answers it; era evaluate makes that call against the environment the student built, and grades their submission against what comes back. No model judges the answers, so the same submission always gets the same score.

title: Week 3 - support triage
day: day1                       # the day of data the answers are about
questions:
  - id: open_tickets
    ask: How many Zendesk tickets are open right now?
    truth:
      system: zendesk           # which system holds the answer
      read: count_tickets       # a read-only tool of that system
      arguments: {status: open}
    grade: number
  - id: pending_tickets
    ask: List the ids of every pending ticket.
    truth:
      system: zendesk
      read: list_tickets
      arguments: {status: pending, limit: 100}
      pick: "[*].id"            # the field to compare
    grade: set                  # partial credit by overlap
    points: 2

$ era evaluate week3.yaml --tenant acme-1 --explain                # each read and its answer
$ era evaluate week3.yaml --tenant acme-1 --submission ada.json   # grade one student
$ era evaluate week3.yaml --class K7QM-3F9P --submissions ./subs  # grade the class

Grading modes: exact (case and spacing ignored), number (with a tolerance), set (F1 overlap, order ignored), sequence (order matters) and contains. then steps - count, sum, min, max, first, last, sorted, unique, lower - shape the answer before it is compared. A read that fails, or that comes back as one page of several, is reported as unreadable rather than graded, and a student whose environment is on a different day than the assignment is flagged, not marked wrong.

The full example assignment and a sample submission are in examples/education.