When you send a message
The judge reads it first
Is this a new task, a follow-up or a correction? Should a message that arrives mid-run interrupt now? Does it need skills and tools, or will a one-line answer do?
◆ Jev multi-step task · full effort · 262 ms
In this turn, the user writes "Jobs with `timeout: "1h30m"` in relay.config.json get killed after about an hour. Find out why, fix it and make sure the tests pass.", and in under a second the judge answers: a multi-step task, full effort. Skills and MCP servers the task doesn't need stay out of the context until it does.
When a tool returns
Output goes through the judge first
A test run or a search can easily return hundreds of lines. The judge decides, chunk by chunk, what matters now; only that enters the context, and the rest is archived behind a pointer, ready when it's needed.
In web pages and MCP results, instructions aimed at the model are withheld before it reads them.
Before acting
Block what's risky, ask only when it matters
Before a command runs, rules flag anything risky, and the judge checks that you asked for it and that it doesn't cross a constraint you stated. There are three permission modes (Full access, Judge mode and Ask first), and you can switch at any time. In Judge mode, the judge lets through what the task clearly needs and asks you about the rest.
When the turn ends
Done means verified
When the model says it's done, mu checks whether anything verified that, whether the turn drifted and whether to go back to a checkpoint.
/goal <condition> keeps the agent working until the condition holds. The main model checks whether it does, with the judge as the fallback; when the same error comes back again and again, the judge decides whether the approach is a dead end.
The hive
Several bees at work, the judge in the middle
When several agents work on one hard problem, splitting up the work is easy; communication is hard. Say nothing, and each walks into the same dead end. Say everything, and every context fills with the others' chatter.
In mu's hive, the judge weighs every finding: is it worth sharing, and could it matter to a given bee? Unless the judge is sure it doesn't, the finding reaches that bee as context, marked as a finding, not an instruction. When a later finding overturns an earlier one, the correction goes to every bee that heard it.
The plain-language board
Working and reporting are different jobs
Models get stronger, and their reports read more and more like notes for a machine. In mu the working model keeps working its own way. Every few tool calls, the judge checks what happened, picks what's really news and hands it to a model chosen for writing plainly, which tells you where things stand.
When a run ends, you see not just "done" but what was actually done.
mu
