Agent Configuration

Agents are AI employees on TeamDay — specialized team members you install, start with, or create for specific roles. This guide covers the settings available when configuring an agent.

Configuration depends on the agent type:

Ready Agents: configure credentials, context, scope, schedule, and approvals. Teamday manages the global prompt, rubric, and upgrades.
Teamday Agents: chat with them and grant local context. Teamday manages their core identity and behavior.
Private Agents: fully editable agents owned by your organization.

Agent Fields

Private Agents have the following configurable fields. Ready Agents and Teamday Agents expose only the fields that are safe for their mutability level.

Field	Required	Description
`name`	Yes	Display name
`systemPrompt`	Yes for Private Agents	System prompt — editable only for organization-owned Private Agents
`role`	No	Role description (default: "Assistant")
`model`	No	AI model ID (default: `claude-sonnet-4-6`)
`category`	No	`marketing`, `finance`, `hr`, `engineering`, `operations`, `general`, `data`
`visibility`	No	`private`, `organization`, `public`, `unlisted` (default: `organization`)
`skillIds`	No	Skills attached to this agent (e.g., `["core:research-assistant"]`)
`mcpInstanceIds`	No	MCP integrations attached to this agent
`subagentIds`	No	Other agents this agent can delegate work to
`allowedTools`	No	Tool allowlist (default: all tools enabled)
`disabledTools`	No	Tools to disable (alternative to allowlist)
`tags`	No	Tags for organization and filtering
`description`	No	Short description (up to 500 chars)
`initialGreeting`	No	First message when a user starts a chat
`image`	No	Avatar image URL
`slug`	No	URL slug for public marketing page
`longDescription`	No	Extended description for marketing page
`useCases`	No	List of use case descriptions
`faq`	No	FAQ items (`[{question, answer}]`)
`integrations`	No	List of integration names
`seo`	No	SEO metadata (`{title, description, keywords}`)
`workspace_access_mode`	No	`all`, `allow`, or `deny` — controls which workspaces the agent can read (default: `all`)
`workspace_access_ids`	No	Workspace IDs for `allow` / `deny` modes — empty list means no extras (`allow`) or no blocks (`deny`)

Equipment

Agents have four types of equipment, each managed via a dedicated tab in the agent detail panel:

Tools

Built-in Claude tools that every agent has access to by default:

Tool	Description
`Read`	Read files from the workspace
`Write`	Write files to the workspace
`Edit`	Make targeted edits to existing files
`Bash`	Run shell commands
`Glob`	Find files by pattern
`Grep`	Search file contents
`WebSearch`	Search the web
`WebFetch`	Fetch and process web pages
`Task`	Launch subagents for complex tasks
`SendMessage`	Communicate between agents
`Skill`	Invoke skills as slash commands

Tools are universal — the same set is available whether the agent is chatting at org level or in a space. Use disabledTools to restrict specific tools, or allowedTools to set an explicit allowlist.

Skills

Reusable capabilities attached to the agent via skillIds. Skills provide domain expertise, workflows, and automation scripts.

{
  "skillIds": ["core:research-assistant", "core:data-analyst"]
}

Skills can also be attached to a space (via skillRefs). When an agent chats in a space, it gets the union of its own skills and the space's skills. When chatting at org level, it only gets its own skills.

See Skills for how to build and manage skills.

Browser Computer is a built-in hosted runner skill for web browsing with a persistent browser profile. Use it when an agent needs screenshots, JavaScript-heavy pages, or authenticated website workflows. See Browser Computer.

MCPs

External integrations attached to the agent via mcpInstanceIds. MCP (Model Context Protocol) servers connect agents to external services like Google Analytics, GitHub, Slack, etc.

{
  "mcpInstanceIds": ["mcp-instance-id-1", "mcp-instance-id-2"]
}

MCPs can also be installed on a space. When an agent chats in a space, it gets the union of its own MCPs and the space's MCPs. At org level, it only gets its own MCPs.

See MCP Servers for available integrations.

Subagents

Other agents this agent can delegate work to, attached via subagentIds. Subagents run isolated tasks and report results back.

{
  "subagentIds": ["agent-id-1", "agent-id-2"]
}

Subagents can also be defined as markdown files in a space's .claude/agents/ directory. The agent discovers both its configured subagents and any filesystem-based subagents in the space.

Equipment Scope: Org Chat vs Space Chat

Context	Equipment Available
Org-level chat	Agent's own equipment only (its tools, skills, MCPs, subagents)
Space chat	Agent's equipment union space's resources (skills, MCPs, subagents from filesystem)

Models

Available Models

TeamDay supports models from multiple providers:

Claude (Anthropic):

Model ID	Name	Description	Pricing (per 1M tokens)
`claude-sonnet-4-6`	Claude Sonnet 4.6	Best value — fast, smart, great for coding	$3 in / $15 out
`claude-opus-4-6`	Claude Opus 4.6	Most capable — complex reasoning	$5 in / $25 out
`claude-haiku-4-5-20251001`	Claude Haiku 4.5	Fastest — simple tasks	$1 in / $5 out

Gemini (Google):

Model ID	Name	Pricing (per 1M tokens)
`gemini-3-flash-preview`	Gemini 3 Flash Preview	$0.50 in / $3 out
`gemini-3.1-flash-lite-preview`	Gemini 3.1 Flash Lite Preview	$0.25 in / $1.50 out
`gemini-2.5-flash`	Gemini 2.5 Flash	$0.30 in / $2.50 out

Teamday only lists Gemini CLI models that have passed a local headless CLI smoke for the current auth path. Antigravity CLI is exposed separately as antigravity-default because agy controls model choice through the connected Google account and does not currently expose a stable --model flag.

OpenAI:

Model ID	Name	Pricing (per 1M tokens)
`gpt-5.3-codex-2026-02-05`	GPT-5.3 Codex (coming soon)	$1.75 in / $14 out
`gpt-5.1-2025-11-13`	GPT-5.1 (coming soon)	$1.25 in / $10 out
`gpt-5-mini`	GPT-5 Mini (coming soon)	$0.25 in / $2 out

Qwen Code (Alibaba Cloud):

Model ID	Name	Description	Pricing
`qwen3.6-plus`	Qwen 3.6 Plus	Flagship coding model	Uses connected Qwen Code account
`qwen3.5-plus`	Qwen 3.5 Plus	Default model for agentic coding and repo work	Uses connected Qwen Code account

Qwen Code models require a connected Qwen Code provider account (Alibaba Cloud Model Studio or an OpenAI-compatible endpoint). Configure it in Settings → Connected Providers → Qwen Code.

Setting the Model

CLI:

teamday agents create --name "Fast Bot" --model claude-haiku-4-5-20251001 --system-prompt "..."

API:

{
  "name": "Fast Bot",
  "model": "claude-haiku-4-5-20251001",
  "systemPrompt": "..."
}

Visibility

Visibility controls who can discover your agent. It does not control where the agent is used — that's determined by attachment (adding the agent to a space or starting a chat).

Level	Description
`private`	Only the creator can see and use
`organization`	All members of your organization (default)
`public`	Listed on the TeamDay `/team` page — anyone can use
`unlisted`	Accessible via direct link, not listed publicly

Public agents get a marketing page at teamday.ai/team/{slug} with their description, FAQ, use cases, and a "Add to Team" button.

System Prompt

For Private Agents, the systemPrompt field is the core of your agent's behavior. Ready Agents and Teamday Agents have Teamday-managed core prompts; configure their local context and credentials instead of rewriting their identity.

Key tips:

Start with identity: "You are [Name], a [role]"
Define behavior: what the agent should do and how
Set constraints: what the agent should not do
Be specific: vague prompts produce vague results

Work Access

Every agent run is anchored to an active workspace — the workspace where the run was started. Writes always go to the active workspace. Reads can optionally span additional workspaces.

Use Work access (Agent Detail → Work access) to control which workspaces an agent can read files from during a run.

Modes

Mode	Label	Description
`all`	All workspaces	Reads every shared workspace (default).
`allow`	Only selected	Reads only the workspaces you explicitly select.
`deny`	All except	Reads all shared workspaces except the ones you block.

Read / Write Asymmetry

Direction	Scope
Runs in	Active workspace (the workspace where the run started)
Reads	Controlled by the mode above
Writes to	Active workspace only — cross-workspace writes require an explicit destination path

This asymmetry is intentional: agents can pull context from multiple sources but always commit work to a single, predictable location.

Edge Cases

allow with no workspaces selected — the agent reads no additional sources ("No extra sources"). This is the most restrictive setting.
deny with no workspaces selected — equivalent to all; no workspaces are blocked.
Workspace access controls affect file reads during a run. They do not restrict which workspaces an agent can be assigned to or chat from.

When to Use Workspace Restrictions

Restrict access when:

Your organization has multiple teams and one team's agent should not read another team's data.
You are running an agent with elevated capabilities (e.g., an admin agent) and want to minimize its blast radius.
You are setting up a customer-facing agent that should only see the customer's workspace.

Leave all (the default) when the agent is designed to synthesize information across your entire organization.

Configuring Work Access

Open Agent Detail for any Private Agent.
Scroll to the Work access section.
Select a mode: All workspaces, Only selected, or All except.
For allow or deny modes, check the specific workspaces to include or exclude.
Save. The new policy applies to all subsequent runs.

Platform MCP Tools

TeamDay provides built-in MCP tools that all agents can use:

Tool	Purpose
`mcp__teamday-media__MediaGeneration`	Generate images and videos
`mcp__teamday-browser-computer__browser_open`	Open and control a persistent hosted browser profile
`mcp__teamday-admin__TeamdayAdmin`	Manage platform resources (spaces, agents, skills)
`mcp__teamday-ui__UICommand`	Control the chat interface (notifications, modals, handoffs)

See Platform Tools for detailed documentation.

Updating Configuration

Private Agent fields can be updated after creation. Ready Agents and Teamday Agents reject protected prompt, identity, rubric, and tool-policy edits.

CLI:

teamday agents update <private-agent-id> --model claude-opus-4-6 --system-prompt "New instructions..."
teamday agents update <id> --skills "core:research-assistant,new-skill"

API:

curl -X PATCH "https://cc.teamday.ai/api/v1/agents/<id>" \
  -H "Authorization: Bearer $TEAMDAY_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"model": "claude-opus-4-6"}'

Only provided fields are updated. Omitted fields remain unchanged. Protected Teamday-managed fields return an authorization error instead of mutating the agent product.

Next Steps

Prompts & Instructions — Write better system prompts
Tools — Deep dive into available tools
Browser Computer — Give agents a persistent hosted browser
Skills — Build and install reusable capabilities
Spaces — Configure workspaces for your agents