How a run works

A run is one piece of work by one agent. A Flow, Apex, or a person on a record page starts it. Pluto Core then calls the model, runs the tools the model asks for, and stores every step on the run.

The steps

sequenceDiagram
    participant S as Flow or person
    participant C as Pluto Core
    participant M as Model
    participant T as Tool
    participant A as Approver
    S->>C: Start Agent Run
    C->>C: Check the agent is shared with the starter
    C->>M: Instructions, context, and input
    M-->>C: Call tool update_case
    C->>A: Approval, if the tool needs one
    A-->>C: Approve
    C->>T: Run the tool as the run's user
    T-->>C: Result
    C->>M: Tool result
    M-->>C: Final answer
    C-->>S: Run Succeeded, answer stored

Each model turn runs in its own Queueable job. A tool call whose tool type is Stays in This Org runs in the same job as the model turn or Reads Outside Systems call before it. Every other tool call gets a job of its own. If a job fails or the org restarts it, Core’s recovery schedule picks the run up again within about 10 minutes. A step can run twice without doing its work twice.

Who the run acts as

The agent’s Agent Execution Identity decides this.

Agent Execution Identity	The run acts as	Can wait for approval?	Can run from a schedule?
Set	The identity’s execution user, a dedicated Salesforce user. A run a person starts is also limited to what that person can see.	Yes	Yes
Blank	Whoever starts the run.	No	No

Use an identity for approvals, background work, and runs that must survive a long wait. Leave it blank for a simple helper that only does what the signed-in person could do anyway. See Create an execution identity.

Run states

stateDiagram-v2
    [*] --> Running
    Running --> Waiting: approval, reply, or outside answer
    Waiting --> Running: answer arrives
    Running --> Paused: limit reached or a step needs a person
    Paused --> Running: operator rescues
    Running --> Succeeded
    Running --> Failed
    Running --> Cancelled
    Waiting --> Cancelled
    Paused --> Cancelled
    Succeeded --> [*]
    Failed --> [*]
    Cancelled --> [*]

Status	What it means	What you do
Running	A step is running or queued.	Nothing.
Waiting	The run waits for an approval, a person’s reply, or an outside system. It ends on its own when the wait expires.	Nothing, or cancel it.
Paused	The run stopped and needs a person. The error code says why.	See Fix a stuck run.
Succeeded	The agent gave its final answer.	Read the answer.
Failed	The run could not finish.	Read the error, fix the cause, start a new run.
Cancelled	A person or the system cancelled it.	Nothing.

What a run keeps

Every run stores its input, each model request and reply, each tool call with its input and result, each approval, and its final answer. Sensitive values are stored as tokens. The data is cleared after the retention period, 30 days by default. See Retention and erasure.