Graph settings
By the end of this page you will know what every graph-wide setting changes, and which ones cost money when you turn them on.
Prerequisites
- An Agent Graph with at least one agent (Canvas tour).
- Admin or Developer role.
These settings live in the Configuration tab of the left panel and apply to the whole graph, not to any one step.
This graph
A read-out at the top: how many Agents are on the canvas and how many Repeats (loops) you have drawn. A quick check that the graph is the shape you think it is.
What agents share
Every agent already sees the answers from the agents before it. This setting decides whether it also sees what they did.
| Setting | Default | What it does |
|---|---|---|
| Share what earlier agents did | Off | Also shows each agent the actions earlier agents took and how they went — not only their conclusions. |
Turn it on when a later step needs to know how an answer was reached — a tester that should see which files the developer actually changed, for example. Leave it off when the outputs are the whole story.
This adds to every agent’s cost, because the summary goes into every downstream prompt. Turn it on because a step needs it, not by default.
Small talk
| Setting | Default | What it does |
|---|---|---|
| Answer greetings and thank-yous directly | Off | A hello, a thanks or a goodbye gets a short reply instead of running every agent. |
Anything that asks for work still runs the graph — and so does anything uncertain. The check errs towards running the graph, so turning this on cannot cause a real request to be answered with a brush-off.
Worth turning on for any graph that people chat with, because running a five-agent pipeline to answer “thanks” is pure cost.
Conditions & repeats
A connection can carry a condition, and a repeat decides when to stop. When those are answered by an LLM rather than by a value check, this is the model that answers them.
| Setting | Default | What it does |
|---|---|---|
| Which AI answers them | Project default | The model for yes/no checks on conditions and loops. |
A small, fast model is the right choice — these are yes/no questions, not reasoning tasks. Leave it on the project default unless you need a specific model.
Final answer
By default the user gets the last agent’s answer, unchanged.
| Setting | Default | What it does |
|---|---|---|
| How to combine the answers | Empty | Instructions for composing every path’s answer into one reply. |
| Which AI writes it | Project default | Appears once you have entered a prompt. |
A composition prompt might read:
Combine the reviews below into one recommendation. Say which reviewer raised each point, and lead with anything blocking.
Leave it empty when one agent already produces the reply you want. With a prompt set, the answer from the end of each path goes into one reply — and if composing fails, the user still gets the answers as they are.
Composing takes a step of its own. On a graph already at the 50-node limit there is no room for it, and the editor warns that the reply will be the last agent’s output instead.
Limits
Safety limits for one request. Usually fine as they are — you are warned before saving if a graph would go over them.
| Setting | Default | Maximum | What it does |
|---|---|---|---|
| Agents working at once | 8 | 20 | How many independent agents may run together. More at once finishes sooner, but costs more at once. |
| Most agents in one graph | 50 | 50 | How many you can place on the canvas. |
| Most agent runs per request | 100 | 100 | The ceiling for a whole request. A repeat runs the same agents again, so this is what keeps a looping graph from getting expensive. |
Most agent runs per request is the one that matters when you use loops. See Loops for how the total is worked out and what to do when a graph exceeds it.
Save your changes
Click Save in the header. The preview runs the saved graph.
Common issues
- Costs rose after a change — Check Share what earlier agents did. It adds context to every downstream prompt.
- Greetings run the whole graph — Answer greetings and thank-yous directly is off by default.
- The final reply is only the last agent’s answer — No composition prompt is set, or the graph is at the node limit and had no room for a composing step.
- Save is blocked over agent runs — A loop pushes the worst case above Most agent runs per request. See Loops.