How to Configure Delegation

🔧 Tools · 📋 Prompt

The delegation provider gives the agent a delegate tool for routing tasks to named specialized agents. Each delegation source contributes a set of available agents the LLM can choose from.

List and Enable Delegation Sources

za agents capabilities delegation-sources list
za agents capabilities delegation-sources configure <source-name>

Configure Provider Settings

za agents capabilities delegation-sources settings

Provider Configuration

Setting	Type	Default	Description
enabled	bool	false	Enable the delegation provider
verbose	bool	true	Include detailed behavioral guidance in tool description and system prompt
include_in_prompt	bool	true	Inject delegation instructions into the system prompt
extra_instructions	str or null	null	Append additional delegation guidance
tool_description	str or null	null	Fully replace the default tool description
allow_llm_selection	bool	false	Let the agent pick the target agent's LLM by name, within its allowed names
sub_agent_approvals	ask, decline, run	ask	What a delegated agent does with a tool call that needs the user's approval, see Approvals in delegated agents

The delegate Tool

When enabled and agents are available, the provider exposes a single tool:

Tool	Description
delegate	Route a task to a named specialized agent. The target agent runs in a fresh context and returns a text result.

The LLM sees an enum of available agent names and picks the right one based on the user's request.

Parameter	Type	Description
agent	string (enum)	The name of the agent to delegate to
description	string	A short (3–5 word) label for the task
prompt	string	A detailed, self-contained task description
llm_configuration_name	string (optional)	Named LLM configuration for the target agent; present only when allow_llm_selection is on

The target agent runs in isolated context: it receives the task prompt and what the user attached to the chat (documents, tags, the file open next to the Office panel), not the parent conversation. It executes with its own tools and returns a single text result. The user cannot see the delegated agent's output directly; the calling agent must relay it.

---

Delegation Sources

Source	CLI name	Description
In-Memory Agents	in-memory-agents	Static list of named agents defined in configuration

---

In-Memory Agents

A static list of named agents defined directly in configuration.

za agents capabilities delegation-sources configure in-memory-agents

Configuration

Setting	Type	Default	Description
agents	list[DelegableAgent]	[]	The list of agents available for delegation

Each agent entry has:

Field	Type	Description
agent_identifier	str	The identifier used to create the agent
name	str	Short display name shown to the LLM in the enum
description	str	One-line description of the agent's specialization
bot_params	dict or null	Runtime bot_params passed to the agent factory
allow_llm_selection	bool or null	Per-agent override of the provider's allow_llm_selection
sub_agent_approvals	ask, decline, run or null	Per-agent setting, which can only be stricter than the provider's

Write descriptions that include trigger phrases — the LLM uses them to decide when to delegate:

{
  "agent_delegation_provider_configuration": {
    "enabled": true
  },
  "in_memory_delegable_agents_source_configuration": {
    "enabled": true,
    "agents": [
      {
        "agent_identifier": "research-agent",
        "name": "research_agent",
        "description": "Performs deep literature research with citations. Use for 'find papers about...', 'literature review', or 'what does the research say about...'"
      },
      {
        "agent_identifier": "data-analyst",
        "name": "data_analyst",
        "description": "Analyzes datasets, creates charts, and computes statistics. Use for 'analyze this CSV', 'plot the trend', or 'summarize this data'."
      }
    ]
  }
}

---

How Delegation Works

• The LLM sees the enum of available agents with their descriptions and picks the right one based on the user's request.
• Delegated agents run in isolated context: they receive the task prompt and what the user attached to the chat, not the parent conversation.
• By default a delegated agent gets no delegate tool, and delegate and task share one budget per turn, counted across every agent the turn starts: see Sub-Agent Limits.
• Protected parameters (tenant, auth headers, index configuration) are carried through automatically.

Approvals in delegated agents

A delegated agent's tool call can need the user's approval (see Tool rules). sub_agent_approvals decides what happens:

Value	Effect
ask (default)	The approval card appears in the user's chat and the delegated agent waits for the decision
decline	The call is declined and the delegated agent is told no approval is available
run	The call runs without asking

Order of strictness is run, ask, decline. A delegation can only tighten the setting: a per-agent value looser than the provider's is ignored, and an agent a delegated agent delegates to is at least as strict as its parent. sub_agent_approvals comes from the agent setup only; a request's bot_params cannot set it. The task tool has the same setting.

Looking for anonymous sub-agents?

The task tool (spawning anonymous sub-agents for parallel work) comes from the sub-agent tool source, not the delegation provider. See Tools — Sub-Agent.