Sub-agents & Delegation
How the agent spawns child agents for parallel work
For complex tasks with multiple independent parts, the Astonish agent can delegate work to sub-agents that run in parallel. This speeds up multi-step workflows significantly.
How delegation works
Section titled “How delegation works”The delegate_tasks tool spawns up to 10 sub-agents simultaneously. Each sub-agent receives:
- Its own isolated session — linked to the parent session for traceability
- Read-only access to memory — sub-agents can reference stored knowledge but do not write to it
- Filtered tool access — only the tools needed for the specific sub-task
- A timeout — 5 minutes by default, configurable
The agent decides autonomously when to delegate. Typically this happens when a task has 3 or more independent parts that each require multiple tool calls.
Sub-agent results are collected and synthesized by the parent agent into a unified response.
The OpenCode sub-agent
Section titled “The OpenCode sub-agent”The opencode tool is a specialized coding sub-agent designed for complex software engineering tasks. The parent agent can delegate implementation, refactoring, or debugging work to it.
Viewing sub-agent work
Section titled “Viewing sub-agent work”To see sub-agent traces inlined with the parent session:
astonish sessions show <id> --recursiveEach sub-agent session is also accessible individually through the standard astonish sessions show command.
Configuration
Section titled “Configuration”These settings can be adjusted through Studio Settings > Sub-agents, or in config.yaml:
sub_agents: enabled: true # enable/disable delegation (default: true) max_depth: 2 # maximum nesting depth (default: 2) max_concurrent: 5 # max parallel sub-agents (default: 5) task_timeout_sec: 300 # per-task timeout in seconds (default: 300)