An agent is the stable product identity. A deployment is an immutable version of that agent’s behavior and sandbox configuration.

​
Agent identity

FieldPurpose
NameHuman-readable label
SlugStable Runtime API model identifier
DescriptionPurpose and operational context
IconOptional visual identifier
Active stateControls whether the agent is callable
Active deploymentImmutable version used by new runs

​
Deployment contents

A deployment pins:
  • model and thinking-level defaults;
  • system instructions;
  • skills and prompt templates;
  • hosted extensions and active tools;
  • runtime environment metadata;
  • managed image configuration;
  • region and machine configuration;
  • workspace upgrade policy;
  • source and manifest integrity evidence.

​
Callability

An agent can appear in GET /api/v1/models when it is active and belongs to the API key’s workspace. Runtime execution also requires:
  • a ready active deployment;
  • valid provider configuration;
  • runtime-ready billing;
  • a key with the required scope.

​
Deployment lifecycle

queued
  -> source uploaded
  -> image build
  -> runtime preparation
  -> ready
  -> active for future runs
A failed deployment does not replace the active ready version. Activating an older ready deployment is a rollback for future runs. Existing durable runs follow their own workspace upgrade policy before changing deployment.

​
Agent source

Agent behavior belongs in the project: 
agent/settings.json
agent/system.md
agent/skills/**
agent/prompts/**
agent/extensions/**
salambo.yaml controls deployment and runtime policy. It is not a second prompt or agent-behavior format.