Sub-Agents and LLM Tasks

Triggerfish agents can delegate work to sub-agents and run isolated LLM prompts. This enables parallel work, focused reasoning, and multi-agent task decomposition.

Tools

subagent

Spawn a sub-agent for an autonomous multi-step task. The sub-agent gets its own conversation context and can use tools independently. Returns the final result when complete.

Parameter	Type	Required	Description
task	string	yes	What the sub-agent should accomplish
tools	string	no	Comma-separated tool whitelist (default: read-only tools)

Default tools: Sub-agents start with read-only tools (read_file, list_directory, search_files, run_command). Specify additional tools explicitly if the sub-agent needs write access.

Example uses:

• Research a topic while the main agent continues other work
• Explore a codebase in parallel from multiple angles (this is what the explore tool does internally)
• Delegate a self-contained implementation task

llm_task

Run a one-shot LLM prompt for isolated reasoning. The prompt runs in a separate context and does not pollute the main conversation history.

Parameter	Type	Required	Description
prompt	string	yes	The prompt to send
system	string	no	Optional system prompt
model	string	no	Optional model/provider name override

Example uses:

• Summarize a long document without filling the main context
• Classify or extract data from structured text
• Get a second opinion on an approach
• Run a prompt against a different model than the primary

agents_list

List configured LLM providers and agents. Takes no parameters.

Returns information about available providers, their models, and configuration status.

How Sub-Agents Work

When the agent calls subagent, Triggerfish:

1. Creates a new orchestrator instance with its own conversation context
2. Provides the sub-agent with the specified tools (defaulting to read-only)
3. Sends the task as the initial user message
4. The sub-agent runs autonomously -- calling tools, processing results, iterating
5. When the sub-agent produces a final response, it is returned to the parent agent

Sub-agents inherit the parent session's taint level and classification constraints. They cannot escalate beyond the parent's ceiling.

When to Use Each

Tool	Use When
subagent	Multi-step task requiring tool use and iteration
llm_task	Single-shot reasoning, summarization, or classification
explore	Codebase understanding (uses sub-agents internally)

The explore tool is built on top of subagent -- it spawns 2-6

parallel sub-agents depending on the depth level. If you need structured codebase exploration, use explore directly rather than manually spawning sub-agents. :::

Sub-Agents vs Agent Teams

Sub-agents are fire-and-forget: the parent waits for a single result. Agent Teams are persistent groups of collaborating agents with distinct roles, a lead coordinator, and inter-member communication. Use sub-agents for focused single-step delegation. Use teams when the task benefits from multiple specialized perspectives iterating on each other's work.