Create an agent
Build an agent in Agent Studio: start blank, from a template or with Solutions Engineer, then choose its model, vault, tools, skills, memory and guardrails.
An agent is a configured assistant you name in a thread and put to work in your workspace. You build agents in Agent Studio, the part of SPARK that holds everything an agent is made from: its model, vault, tools, skills, memory and guardrails.
Think of an agent as a teammate with a role: it sits beside people in the same channels and threads, and you hand it a piece of work as you would a colleague. A workspace can have several, each with its own role, such as a NAV Oversight Analyst and a Reconciliation Break Investigator, and they work with your team rather than replace it.
Every workspace also comes with built-in agents: Spark, which answers any thread that names no other agent, Solutions Engineer and Coding Assistant. See Built-in agents.
Open Agent Studio
Agent Studio has three ways in:
- On the Ecosystem tab of the sidebar (⌘⇧E on a Mac, Ctrl+Shift+E elsewhere), click Studio. It opens on the Agents page. Rest the pointer on Studio to see every Agent Studio page beside it: Agents, Models, Skills, Tools, Memory, Connectors, Vaults and Guardrails.
- On the Ecosystem page, the Agent Studio row has a card for each of those pages.
- Open the Settings tab (⌘⇧S on a Mac, Ctrl+Shift+S elsewhere). It always opens on Agent Studio's Agents page, and the Agent Studio group at the top of the settings sidebar lists the same pages.
The Agents page
The Agents page, AI agents, each with its own model, instructions and tools., lists the workspace's agents as cards, two across. Each card shows the agent's lens in its colour, its name and two lines on what it does, then its model and how many tools and skills it has. Rest the pointer on the tool or skill count to list them. The model chip is followed by where the model runs, such as the Google Cloud and EU marks after NGT Auto.
The magnifier at the top widens into a search box that filters by name. The archive button shows archived agents (Show archived), and New makes one.
Click a card to open the agent. Its ⋯ menu, which shows when you rest the pointer on the card, holds Edit agent, Rename, New thread, Archive and Delete. An archived agent offers Restore and Delete instead.
Below your own agents, a folded Built-in agents section lists the agents that come with SPARK. See Built-in agents.
Rename an agent
Choose Rename from a card's ⋯ menu to change the agent's name and description in a dialog, without opening the editor. On an open agent, click its name in the page header, or use Rename in the header's ⋯ menu. A name can have spaces, such as NAV Oversight Analyst: it is how people tag the agent in a thread.
Create an agent
Click New to open a blank agent. The arrow beside it opens a menu with three ways to start:
- Blank agent: a new agent that works as it is, the same as clicking New. It starts with short, generic instructions and every tool of a thread, so it can list, read, post, reply and delete in the workspace's channels and threads; change or untick any of them before you save.
- From template: a ready-made agent to adjust. See Agent templates.
- Ask Solutions Engineer: starts a thread with the built-in Solutions Engineer, which asks about your processes and proposes a setup. This item shows only while Solutions Engineer is switched on for the workspace.
You can also ask an agent to make one: Spark, or any agent you give the tools to, can create an agent from scratch or from a template. See Tools.
The agent page
An agent's page is one scrolling form of sections on the left, with a live Preview beside it on the right. On a narrow window the two stack.
Click a section's title to fold it open or shut. A folded section shows a summary of what it holds next to its title, such as 3 tools or the vault's name. The question-mark icon after a title explains the section in one sentence. The fold button in the page header folds or unfolds every section at once.
Sections that pick from a list (Tools, Skills and Guardrails) have a + in their header that opens a picker dialog, and an Add button while they are empty. They show their first ten picks as chips, then how many more (+3 More...). Most sections also have a gear that opens the matching Agent Studio page, such as Manage models or Manage tools, so you can adjust it without losing your unsaved edits.
The sections, in order:
| Section | What it does |
|---|---|
| Name | The agent's display name, up to 50 characters. Click the lens in the header to change its colour. |
| Description | Optional. A short summary of the agent's purpose, shown on its card. |
| Model | The AI model the agent thinks and answers with. NGT Auto picks a model for each message. See Models. |
| Vault | The credentials the agent uses: your own model keys and the accounts of its connected apps. See Vaults. |
| Tools | The actions the agent can take, in SPARK and in the apps you connect. See Tools. |
| Skills | Reusable instructions the agent loads when its task calls for them. See Skills. |
| Memory | The memory store the agent saves to and recalls from in later threads. See Memory. |
| Guardrails | Checks that run on every model and tool call the agent makes. See Guardrails. |
| Instructions | What the agent is told in every thread: its role, its rules and how it answers. |
| Advanced | Folded by default. Temperature, Thinking and Max steps. |
The Vault sits right under Model because the vault decides which models you can pick: models that run on your own provider keys only appear once you choose a vault that holds the key. If you switch to a vault that cannot run the selected model, SPARK moves the agent to NGT Auto and tells you.
Write the instructions
Type the agent's instructions in the Instructions section, or click the expand icon (Expand instructions) to edit them in a larger window. With the agent's name and description, the instructions define its role, much as a title, a remit and a job description do for a colleague. Clear instructions are the single biggest lever on quality.
A blank agent starts with generic ones, so you need to write your own only when you want to: Help the people you work with reach their goals. Use your skills, tools and memory to find what you need and get it done, within this workspace's guardrails. Ask when something is unclear. An agent made from a template starts with the template's instead.
Advanced settings
Open Advanced to tune how the agent's model runs:
- Temperature: a slider from 0 to 1. Lower is more focused, higher is more varied.
- Thinking: Low, Medium or High reasoning effort. It is disabled, with a tooltip, when the model does not support thinking.
- Max steps: the most calls the agent can make to the model to answer one message, from 1 to 50. The default is 20. One call can run several tools. When it runs out, the agent stops and sums up what it found.
Preview, then save
Once the agent has a name, a model and instructions, the Preview unlocks: until then it reads Fill in the name, model, and instructions to preview your agent. Type a prompt in Enter a prompt to preview... to try the agent before you commit. Before the first prompt, the Preview shows the agent's model and how many tools and skills it has. The preview allows a limited number of messages; when you reach them, Preview limit reached asks you to save the agent to continue (Save the agent to continue chatting.).
A preview reads the agent's memory store but never writes to it, so trying a prompt does not leave memories behind.
Click Create agent to save a new agent, or Save on an existing one. The button stays disabled until the name, model and instructions are filled in.
Who made it
A saved agent's Preview shows where it came from: who created it (By a name), when it was Created and Updated, and how many revisions it has, beside its model and its tool and skill counts. Agents created before this was recorded show no author.