Mercury beta.22: the crew moves into the chat
I’ve been building beta.22 with a fleet of agents on different models, using Jev alongside them to help decide what needs a closer look.
What I’m using Jev for
In the last Jev post, I said I wanted to try it on Mercury’s development. I’ve now been using it on failed tests and on the writing for the project.
Jev’s job is to judge the evidence it is given. A session sends it a log or a piece of text through JevEval, asks specific questions, and gets numbers back. The coding agents do the writing, and I make the design decisions.
Beta.22 also adds another way to connect to Jev, with its own key, usage meter and spending cap. It stays off until you enable it, and the agents can use it too.
When a test fails in a test run or pool, the session gives Jev the log and asks two questions: was the process killed, and was it starved of resources?
I use 0.6 as the cutoff for a rerun. If Jev scores either answer that high, the failed check runs again on its own. Otherwise, the log goes to a person to investigate. The test results still decide whether the code passes.
I’ve also used it to check the README one block at a time, looking for jargon and awkward phrasing. I kept the factual corrections even when the scores were poor. Jev helps me find passages to look at; I still choose the words.
Keeping track of the agents
With a fleet working inside Mercury, I want to be able to see what each agent is doing and speak to the right one. Beta.22 adds a crew view inside the main chat. Opening an agent shows its transcript in the centre, and you can direct the composer to it.
The CREW box keeps the agents the session has used, including the ones that have finished or stopped, until you clear them. Their state stays visible, so finishing a piece of work doesn’t make its row disappear.
The composer now keeps its text when you change views, switch the target or open a pop-up. Messages sent to an agent leave the composer straight away and come back if they are refused. The hints tell you who will receive the next message.
Interrupting the lead ends its turn while the crew keeps working. The next message can start a new turn immediately, and held notices arrive with it. Pausing a workflow now pauses that workflow’s agents alone.
The rest of the update
I kept finding moments when I wanted to speak my thoughts instead of typing them. I wanted to make that easy to use repeatedly while working, so beta.22 smooths out the dictation controls. With dictation enabled, hold space for a second to record, then release it to put the words in the composer.
If a background session asks permission while nobody is connected to answer, it now gets a denial with a reason and can continue its turn. That keeps the refusal attached to the action it blocks, instead of aborting the whole turn.
The project notepad and /note are also being removed. WORKBENCH takes the vacated card slot, and existing saved notes are left untouched.
Beta.22 also includes fixes for account usage displays, expired sign-ins, interrupted tool calls and Windows path handling. The full release notes cover those individually.