hangsiahong/dispatching-parallel-agents

Parallel Agent Dispatch (tmux-Integrated)

Overview

Speed up development by offloading independent tasks to parallel Worker Agents (Gemini-3-Flash) while you (GLM-5.1) coordinate from the Orchestrator pane.

The Workflow

1. Identify Independent Tasks

Group your implementation plan into chunks that don't touch the same files.

• Task A: src/auth/*.ts
• Task B: src/models/*.ts
• Task C: src/routes/*.ts

2. Distill Context (CRITICAL)

Before spawning agents, you MUST extract the current state so the workers aren't lost. Use the Session Distiller:

• Prompt the Distiller to create a "Context Checkpoint" for the specific task.
• Ensure the distillation defines Scope & Anti-Goals (what not to touch) and Verification steps.

3. Dispatch with observability

For each task, use the spawn_agent tool. In the message to the agent, pass the Context Checkpoint generated in Step 2, assign them a Worker ID, and tell them to log their progress.

Example Dispatch Command:

# 1. Create the pane in tmux (run this yourself)
bash ~/.codex/superpowers/skills/tmux-workflow/scripts/tmux-manager.sh add-worker "worker-1" "Auth Implementation"

# 2. Spawn the agent (use spawn_agent tool)
# Prompt: "You are a worker agent. 
# Here is your Context Checkpoint: [Paste Distilled Context Here].
# Log progress using:
# bash ~/.codex/superpowers/skills/dispatching-parallel-agents/scripts/worker-log.sh worker-1 progress 'Message'"

4. Monitor

Watch the tmux panes. Each pane tails its own log file (/tmp/codex-logs/worker-id.log). You'll see real-time updates as the sub-agents work.

5. Collect & Merge

Once agents finish, they will report back in the main session. Review their work, run tests, and commit.

Why this works

• GLM-5.1 is the Brain: Handles the reasoning and planning.
• Session Distiller is the Filter: Prevents context bloat and enforces surgical boundaries.
• Gemini-3-Flash is the Hands: Handles the coding and testing in parallel.
• tmux is the Eyes: Provides a real-time dashboard of all active workers.

Best Practices

• Isolation: Ensure agents don't edit the same files at the same time.
• Small Batches: Give agents specific, bounded tasks.
• Strict Scope: Always use Distiller to give agents Anti-Goals (e.g., "Do NOT refactor the router").