Skip to content
Taikon Docs

Agents

Agents (/admin/agents) covers the full lifecycle of an agent: its identity, which LLM provider it prefers, what it’s allowed to do, how it behaves, who can reach it, and its API keys for external invocation. See Agents & tasks for what an agent is from a user’s point of view.

Creating or editing an agent

Section titled “Creating or editing an agent”

“New Agent” opens a modal organized into sections.

Identity - Name (required) and Model. Leaving Model blank means the agent uses the tenant’s default model instead of a specific one.

Provider - Preferred provider. Leaving this blank means the agent inherits whichever provider the tenant defaults to, rather than a fixed choice.

Capabilities - Skills and MCP servers, each as a searchable checklist with Select All and Select None. These are the tools and MCP-backed capabilities this agent can use; a capability not checked here is unreachable by this agent no matter what any user’s own grants allow.

Behavior:

  • Approval mode - needs_approval or auto. auto lets this agent’s tool calls proceed without a human approval step. This is still subject to the org-wide automation kill switch on Settings - governance: if that switch is off, no agent can run in auto mode regardless of its own setting.
  • Allow invocation via external surface - a toggle exposing this agent to the external invoke API, separate from ordinary chat access.
  • Personality - a free-text instructions field, capped at 4096 bytes with a live counter.

Access (edit-only) - an “Assign to” input taking a user:email or group:name, plus a button to add it. The assignment list below shows everyone and every group with direct access, each with a per-row Revoke.

API keys (edit-only) - minting a key shows the raw key exactly once, in a banner with a Copy button; it cannot be retrieved again after you close it, so copy it before dismissing the banner. The list below shows existing keys by prefix and label only, each with Revoke. The mint form takes an optional label to help you tell keys apart later.

The agents table

Section titled “The agents table”

Each row shows the agent’s name, model, an approval-mode badge (with an additional “ext” badge if external invocation is enabled), skill and MCP counts, how many users/groups it’s assigned to, and Edit/Delete actions.