Trying a change
Try a change on messages the agent already answered before it goes live, or talk to it in its own chat.
Part of the crew rollout: Agents shows as Soon in the app until it reaches your account. See Hiring your crew.
There are two ways to see what an agent will do: try a change on messages it already answered, before you apply it, or talk to it.
Try it on past messages
A change to what the agent knows, its instructions, tools or skills offers a button on its card: Try it on 3 past messages · up to $1.00. The price is what the last test of this change cost, or the most it may cost.
It picks the messages the change is about, up to five. Otherwise it takes the agent's latest failed or low-graded runs, plus one that went well, so a fix can't quietly break what worked.
Each message runs again on the changed agent, as a test run:
- It reads its memory, with the change's memory in it, and changes none of it: no memory written, no skills learned, and its outcome doesn't count for or against a skill.
- Replies, emails, posts, payments and code are recorded, not carried out. The result says what it did instead: "Would have: …".
- It opens a web page of its choosing only when the message was yours or a schedule's, so a stranger's message can't make a test fetch a page.
- If it would have asked a question, the result says "Asked instead: …".
The card sums it up, for example "Tried on 3 past messages · 2 better, 1 same · $0.14". Open it to read each message's before and after, each graded.
"Before" is the answer it gave then, with real tool results; "after" is the test. So the two can differ in more than the change. Co Builder can also run the live version the same way on each message, so the change is compared like with like; that doubles the runs and the cost.
Change the card after testing it and it says "Tested an earlier version of this change", and the button comes back.
Ask Co Builder to test
"Try it on the last three Telegram messages", or, with no change open, "does it still fail?" Co Builder can test on runs you name, on messages you write, or on its own pick. Each test is billed, so it asks you first and says what it will cost.
Talk to it
Open the agent's conversation in the workspace and message it. That is a real run: it does what it's built to do, with its real tools, and learns from it the way your own runs do. Each reply has Change how it does this, which takes that reply to Co Builder.
On the memory map, the agent and the task that started it light up while it runs. Its Runs list every run, with how long it took, what it cost, and what it did.
Say hi on a chat app
Before Who can talk, the connect dialog has Say hi: you send the new bot a message with a code, and the agent answers you there. See Chat apps, schedules & going live.