One Question, More Than One Agent

An agentic question used to force a choice before the candidate ever saw it: our own agent, or one outside command-line agent (a CLI), picked once by the person building the assessment. That was a simplification we did not need to keep.

What changed

A question can now grant a candidate more than one coding agent at the same time, our own in-browser agent alongside outside CLIs. When more than one is granted, the candidate sees a tab per agent and can move between them mid-question. Each tab keeps working while another is open, so switching tabs to check on something does not pause the one you left.

The same choice now works the same way in three places: a hiring assessment, a classroom coding assignment, and the self-serve building product. You learn the picker once instead of relearning a different one for each. In a classroom, a teacher can hand students one specific agent or let them pick their own. When a recruiter asks Assessment Copilot for a question with a command-line agent but names none, it now grants every agent on offer.

Two bugs surfaced and got fixed along the way. A question configured to grant more than one agent could silently lose part of that grant, saving fewer agents than the assessment actually specified. And a batch of library questions listed our own agent alongside both outside ones, but any question that listed our own agent ran with that agent alone, so the outside two never started. Multi-agent support only means something if the agents you configured are the agents that actually show up, so both were treated as real bugs.

Why it matters

Real engineers reach for whichever tool fits the moment and switch without thinking about it, sometimes mid-task. A question that only allows one agent measures how well someone performs inside an unfamiliar constraint, not how they actually work day to day.

Letting a candidate move between agents, and showing a reviewer which one they reached for and how much, is a better proxy for the skill everyone actually cares about: judgment about which tool suits which task. A candidate who tries one agent, decides it is not converging, and switches to another mid-question is showing exactly that judgment.

The report counts the prompts a candidate sent to each granted agent, question by question. An agent they were given and never used shows a zero instead of disappearing, since leaving a tool untouched is part of the picture.

As before, a recruiter sets how much AI help a question allows before deciding which agents are on offer, and a "No AI" question shows plainly that no agent was ever available. Multi-agent choice widens what a candidate can do without loosening the rules a recruiter already set.

Where it points

The tab a candidate is not looking at still needs to be legible when they come back to it. A tab whose agent is still working now lights up while it runs and marks itself once it finishes somewhere off screen. The ongoing project is making multi-agent work feel calm instead of turning one question into three things to babysit at once.