Skip to main content
Everything about how your agent works is a sentence you can say to it. There is no settings page with a temperature slider, because the settings are code in your harness and the agent is the one who edits it. Four asks, in rough order of how often people make them.

​
”Be shorter”

The answer came back in ninety seconds and was two sentences long:
And the harness was untouched. Nothing was edited, nothing was published, and git status in the checkout was clean. The agent read that as an instruction for the conversation and simply followed it — which is right, and which lasts exactly as long as this chat.
This is the trap on this page. A behavior ask, phrased as a request, is usually answered as a request. It works, you see it work, and then a new chat starts and the agent is wordy again.If you want it to stick, say so:
Then it is a file in the harness, a version, and every future conversation.
What the agent edits when you do say that is its own prompt — the file its declaration reads at the start of every run. The edit lands on the very next message in this chat, with nothing published in between, because the platform asks the harness who the agent is at the moment it needs to know. The publish is what carries it to new chats. The same shape covers most of what people mean by “behave differently”:
The starter prompt already leans this way — “Prefer doing over asking: pick a sensible default and go, and come back for a decision only when getting it wrong would cost something that cannot be undone.” If your agent is asking you too much anyway, the useful ask names the case: “stop asking me before you run shell commands” is actionable in a way that “be more autonomous” is not.
Check it by the next answer being shorter. That sounds glib, and it is the real check: this is the one change whose effect you see immediately, in the thing you are already looking at. Undo it by saying so — “go back to how you were answering” — or by rolling back the version, if the change was published and you want it gone everywhere.

​
”Use a cheaper model when we are only chatting”

Two different things wear this sentence, and it is worth knowing which one you want.

​
For this conversation only

The chip in the composer. It reads Harness decides · Default until you touch it.
The Splox model picker open over the composer, showing Model, Effort and Advanced

The composer chip: Model, Effort, and the endpoint under Advanced

  • Model — Harness decides means whatever your agent declares. Pick one and it overrides the harness for this chat.
  • Effort — how hard the model thinks before it answers, on a ladder of Off, Low, Medium, High and Max.
  • Advanced → Endpoint — whose credential answers: the platform’s, or one of your own from Connections.
The choice is per chat, but a new chat starts on the last model you picked, so put it back to Harness decides when you are done experimenting. Nothing about this touches your harness, and nothing about it is permanent. Under the answer is the footer, which says which model actually ran:

​
As a rule, every time

That is a rule, so it goes in the harness. The agent edits the file that answers “which model, this turn” — one of the points of the run loop it can take over — and the rule is asked again on every single turn rather than read once. The starter ships that file doing nothing at all, with the shape written into its own docstring:
A rule like this is worth being concrete about. “Cheaper when chatting” needs a test the code can apply — the turn number, whether the message mentions a file, whether any tool has been called yet — and if you do not supply one the agent will invent one. Ask what test it used.

​
”Check your work before you answer”

Your agent can hand work to another agent and read the answer back in its own code. The starter already declares two: an Assistant that talks to you and an Executor it can hand a self-contained job to. Asking for review makes the second one a reviewer. The other shape is a rule at the point where the loop decides the turn is over: that point can send the model back around with a follow-up instead of letting the answer out. It is the right one when “check it” should apply to every answer rather than to the ones the agent remembers to check. Which you get depends on what you asked for, and it is worth asking which you got, because they cost differently: a sub-agent review is a second run with its own bill; a rule at the end of the turn is another pass of the same one.

​
What it costs, and what it cannot do

Every one of these is a run. A prompt edit is a minute or two; a rule the agent has to write and test is longer. The exposure is the same as any other ask: the previous version is still in the list. A prompt is not a guarantee. “Never delete anything” in a prompt is a strong suggestion, not a lock. If you need it to be a lock, ask for it as a rule at the point where tool calls are approved — that point sees every call by name with its arguments before it runs and can refuse it. Say which one you want: “I want this enforced, not just asked for” is a sentence the agent understands. Some things are per chat and cannot be a rule. The model chip and the effort setting are conversation-level overrides. If you want them every time, they have to be in the harness.

​
How to check a rule actually fired

The answer being different is the first check. The second is that every run keeps a record of which of your rules ran and what each decided:
That is from a real run on the account this page was written from, before any model rule existed. use_default means the file was reached and chose not to decide — which is exactly what a rule that is not firing looks like, and exactly what a rule with a typo in it looks like too.
A rule that raises does not stop the run. The platform’s own answer runs in its place and the turn continues, which is what you want in production and precisely what hides a mistake while somebody is writing one. So “it didn’t work” is a real possibility even when nothing looked wrong, and the way to settle it is to ask: “did your model rule actually run on that turn?” The agent has a tool that reads the run’s journal and answers it point by point.
The third check is the version list: a rule that is not in it was not published, so it is not there in a new chat. See Versions.

​
Undoing

If the agent has become bad enough that asking it to fix itself is not working — it will not stop apologising, or it now refuses ordinary work — do not keep asking. Pin the machine to an older version instead: When it goes wrong.

Which model to reach for

What the models on this platform are good at, and what effort changes.

The reference your agent reads

The ten points of a turn, exactly. Hand your agent this URL when you want a rule at a specific point.