Multi-agent orchestration with delegate-task
Parent agent plans, subagents execute. One clean context per task. How real orchestration actually looks.
⚡ 6min read expert OS · macos · linux v1.0.0
Last updated May 08, 2026 · by xlrd
Prereq
- Hermes Agent 1.0+ or Codex CLI with delegate support
- Anthropic / OpenAI / OVTH key
- A task big enough to split
Steps
01. Sketch the parent prompt
The parent’s job is planning and synthesis, not execution.
markdown
You are the lead agent. Break the task into 3 independent subtasks.
For each, spawn a subagent via delegate_task with:
- a specific deliverable
- the exact files it may touch
- a success check
Then merge the summaries into one PR description. 02. Delegate with scoped briefs
json
{
"tool": "delegate_task",
"args": {
"agent": "worker",
"task": "Refactor src/auth/session.ts to use new token schema. Return diff only.",
"allowed_paths": ["src/auth/**"],
"budget_tokens": 40000
}
} 03. Parallelize where safe
bash
# Three subagents, three independent scopes
delegate-task --task "update payment module" --paths "src/payment/**" &
delegate-task --task "update shipping module" --paths "src/shipping/**" &
delegate-task --task "update notifications" --paths "src/notify/**" &
wait Keep scopes disjoint. Merge conflicts between subagents are a you problem.
04. Synthesize in the parent
Parent reads the three summaries (not the full transcripts), writes the PR description, runs the test suite, opens the PR.
05. Use git as the integration bus
Each subagent commits to its scoped branch. Parent merges them in order. git is the shared memory; no in-memory coordination needed.
Next
- Hermes Agent Telegram orchestration
- Pattern: supervisor + workers + critic
Feedback · anonymous
Was this helpful?