Build an Agent

Create an agent in the Agent Builder, upload one, fork a template, or connect a GitHub repo, and configure every setting

There are three ways to start an agent: build it from scratch in the Agent Builder, fork a template, or connect a GitHub repository that already holds one. All three produce the same thing — a folder whose CLAUDE.md is the agent’s prompt — and use the same runtime settings.

Start from	Button on the Agents page	Where the files live	How you edit
Scratch	Create New	Workspace storage	Agent Builder
Files you already have	Upload	Workspace storage	Agent Builder, or Update files
A GitHub repository	Connect repo	Your repository	Push to the repo
A template	Start from a Template during onboarding	A fork in your GitHub account	Push to the fork

Each workspace keeps agents in one place, chosen when the workspace is created (Context Mode: STORAGE (object store) or GIT (connected repo)). Create New and Upload appear in storage workspaces; Connect repo appears in workspaces on DataGen Cloud. Self-hosted installations always use storage.

---

Create an agent from scratch

1. Open Create New

   Go to Agents and click Create New.

2. Name it

   Agent name becomes the agent’s file root and its URL slug (for example in the Agent API path /api/agents/<name>/v1/messages). It cannot be changed later. Add an optional Description.

3. Pick its runtime

   Under Runtime, choose the Vault, Tool bundle, Default model and Sandbox environment. Leave any of them on the default to use the workspace default. You can change all of them later in the builder.

4. Open the builder

   Click Create & open builder. DataGen creates the agent and opens it in the Agent Builder.

---

The Agent Builder

The Agent Builder lives at /signalgen/agents/<agentId>/build. Open it from Create New, or click Edit on any storage agent’s page.

Column	What it holds
Left	Tabs: Chats (your builder conversations), Config (every setting below), Versions (published versions and rollback)
Center	The Chat with the agent’s sandbox, any files you open, and the Test run tab
Right	Sandbox files: the agent’s working tree, the Changes tab, and Shared files (read-only) for the knowledge bases it can read

The builder keeps one long-lived sandbox per agent. Everything you and the chat change happens in that sandbox’s working tree; nothing reaches the live agent until you publish.

Build with the chat

Ask the chat to write or change the agent. It edits files in the sandbox and you see them update on the right. The empty chat offers starter prompts:

• Write my CLAUDE.md — document what the agent does, its inputs and outputs, and how to work with its files
• Add a skill that… — scaffold a skill for a task you describe
• Review this agent — suggest improvements to its instructions, skills and structure
• Set up a weekly schedule — plan a recurring run for the agent

Edit files directly

Click a file to open it in the center column, edit it, and click Apply (Cmd/Ctrl-S) to write it to the sandbox. Use New to scaffold a New skill, New command or New sub-agent with valid frontmatter, and Upload to add files or a whole folder. Entries in the tree can also be renamed, downloaded, reverted or deleted.

Test before you publish

Click Test run to run the agent on your current draft in its own tab. Starter prompts include Run a smoke test, What can you do? and Run an example task. Click New run to start fresh on the latest draft.

Test runs work even while the agent is disabled, so you can build and test without anyone else reaching it.

Publish

The Changes tab lists every file that differs from the live agent. Click Publish (the button shows the number of changes), add optional Release notes, and confirm. Publishing:

• Syncs the sandbox files to the agent’s storage
• Records a new version, including the agent’s settings, that you can roll back to
• Enables the agent if it was disabled

Click Discard to throw away all unpublished changes, or Revert a single file. See Versions and Logs for rollback.

Restart the sandbox

Vault, tool bundle, harness and model changes apply on your next message. A changed or rebuilt sandbox environment only applies to a fresh sandbox: click Restart sandbox in the header. Unsaved files are checkpointed and restored; anything installed by hand inside the sandbox is lost.

One editor at a time

Only one person can be active in an agent’s builder. If someone else is using it, choose Take over to become the active user, or View only to browse the files and chat history.

---

Agent file layout

my-agent/
├── CLAUDE.md                     # The agent's prompt
├── .claude/
│   ├── skills/<name>/SKILL.md    # Skills the agent loads when relevant
│   ├── commands/<name>.md        # Slash commands
│   └── agents/<name>.md          # Sub-agents the main agent can delegate to
└── scripts/, data/, ...          # Anything else the agent uses

Skills, commands and sub-agents belong to the agent. They are not run on their own: only the agent is enabled, automated and reached through channels.

Skill: .claude/skills/<name>/SKILL.md

---
name: enrich-lead
description: Enrich a company lead with web research. Use when given a company name or domain.
---

# enrich-lead

Given a company name and domain, find:
1. Company size, industry and funding
2. Recent news and product launches
3. Key decision makers and their titles

Return structured JSON with all findings.

A skill folder can hold supporting files next to SKILL.md, such as reference notes or helper scripts.

Command: .claude/commands/<name>.md

---
description: Summarize this week's new leads.
argument-hint: "[region]"
---

Summarize the leads created in the last 7 days for $ARGUMENTS,
grouped by source, and flag anything unusual.

Sub-agent: .claude/agents/<name>.md

---
name: report-writer
description: Delegate here to turn analysis results into a short report.
---

You are a focused sub-agent. You receive analysis results and return
a concise report with a summary, three key findings and next steps.

---

Configuration reference

The builder’s Config tab holds the settings below; click Save config after changing them. The agent page has the same settings spread across its Settings, Tools and Data access tabs, plus a few that only live there. Changes apply from the agent’s next turn.

Identity

Field	Notes
Agent name	Fixed after creation. It is the agent’s file root and URL slug.
Description	What the agent does.

Model and runtime

Field	Where	Notes
Default model	Config, Settings	The model the agent runs on when neither an API call nor an automation names one. Leave it blank for the system default, shown as the placeholder. DataGen checks the model against the provider when you save.
Harness	Settings tab, Harness & model	The harness preset that runs the agent. Default (built-in SDK, stock Anthropic) uses the built-in Claude Agent SDK. If the preset pins a provider model, the agent always runs on that model and the default-model field is hidden. See Harness.
Sandbox environment	Config, Settings	The baked image and packages the agent’s sandbox starts from. Default (standard image) uses the standard image. Appears when sandbox environments are enabled for your account. See Sandbox Environments.

Tools and credentials

Field	Where	Notes
Vault	Config, Settings	The credential set the agent’s tools and code authenticate with. Workspace default uses the workspace’s default vault. See Vault.
Tool bundle	Config, Settings	The allowlist of tools the agent can call, chosen from bundles in the selected vault. Changing the vault resets the bundle. See Tools.
Enable DataGen MCP for this agent	Config, Settings	Attaches the DataGen MCP gateway (searchTools, getToolDetails, executeToolCall) to the agent’s runs. Off by default.
Enable workspace data tools for this agent	Config, Data access	Lets the agent list and read the workspace’s raw files, its own earlier outputs and its connected knowledge bases on demand. Read-only. On by default.
Skills	Tools tab	Install approved skills from the workspace registry. See Skills.

Knowledge and data

Field	Where	Notes
Connected contexts	Config (Knowledge bases), Data access	Knowledge bases the agent reads, mounted read-only at /mnt/context/contexts/<slug>/. Each run reads the live version. A knowledge base attached Read & write also serves as the agent’s memory. Read vs. read & write is set on each knowledge base’s own Agents tab. See Use Context.
Exposure	Data access	Connected only (the default): the sandbox holds exactly the connected knowledge bases; raw workspace files are fetched on demand through the workspace data tools. Whole workspace is a compatibility mode for older agents: once switched to Connected only, it cannot be switched back.
Analytic models	Config, Data access	Published materialized models the agent can query with the datagen-data CLI. Each query reads the live version’s current data. See Use Data.
Data scope	Data access	Which raw workspace data the agent may read: Unrestricted, or Selected folders only with the folders and files you add. A hard boundary: anything outside it is refused. The builder shows a summary with a link to change it.

Behaviour and memory

Field	Where	Notes
Show working details in chat	Config, Settings (Chat display)	On: chat shows tool calls and reasoning beside the answer. Off: answer only. Builder chat and run logs always keep full detail.
Auto-record learnings	Config, Data access	Adds DataGen’s default memory-maintenance guidance to the system prompt, so the agent records what it learned into any knowledge base it can write to. Off by default, and only takes effect with a writable knowledge base. See Memory.
Agents it can call	Data access	Other agents this agent can start a turn on with datagen-assets invoke, as the person it is running for. A called agent cannot call further agents.

Skills

Skills from the workspace registry are installed per agent on the Tools tab:

1. Approve the skill under Skills in the sidebar. See Skills.
2. On the agent’s Tools tab, search for it in the Skills card and install it.

An installed skill is copied into the agent’s own .claude/skills/ and ships with its files on every run. When a newer approved version exists, the card offers to update it.

---

Upload an agent

If you already have an agent folder, upload it instead of building from scratch.

1. Click Upload

   On the Agents page, click Upload.

2. Choose a file or folder

   Pick Single file or Folder, and enter the Agent name. Files are written under <name>/, and the dialog lists exactly where each one lands.

3. Upload

   DataGen creates the agent. Its CLAUDE.md is the prompt.

On the agent page, Update files uploads a new version of the folder, and Export ZIP downloads the agent so you can use it anywhere.

---

Connect a GitHub repository

Use this when your agent already lives in a repository, or you prefer to author it locally in Claude Code and push.

1. Click Connect repo

   On the Agents page, click Connect repo.

2. Install the GitHub App

   Click Install GitHub App and install the DataGen GitHub App on your organization or personal account. You return to the dialog when it finishes.

3. Select the repository

   Choose it under Select a repository, or use Advanced: Enter URL manually for a repository that is not listed. Click Connect.

DataGen creates one agent per repository. The repository’s CLAUDE.md is the prompt, and the skills, commands and sub-agents under .claude/ are listed on the agent page under Skills & commands. Every run loads the sandbox from the repository’s checkout.

To change a repository agent, push to the repository. The runtime settings (vault, tool bundle, model, harness, sandbox environment, knowledge and data) are set on the agent page the same way as for any other agent.

Using Claude Code locally? The DataGen plugin’s /datagen:build-agent command walks you through designing an agent, and /datagen:deploy-agent connects the repository. See Claude Code Plugin.

---

Fork a template

New DataGen Cloud accounts are offered Start from a Template during onboarding. Use Template forks the template’s repository into your GitHub account and connects it, so it behaves like any other repository agent. See Templates.

---

Next

Give it tools

Vaults, tool bundles, MCP servers, custom tools and the DataGen MCP gateway

Trigger it

Run now, schedules and webhooks with automations