Mixture Of Experts

Mixture of Experts (MoE)

Spawn multiple specialized experts in parallel, each analyzing from a unique angle. Aggregate their insights into a comprehensive, multi-dimensional answer.

Foundation: This skill builds on spawn-subagents for the underlying tmux/pi subagent mechanics. Read spawn-subagents for the core patterns before using MoE.

When to Use

• Architecture or design decisions
• Complex code reviews
• Security audits
• Performance optimization
• Debugging tricky issues
• Evaluating trade-offs

Expert Specializations

Core Experts

Domain Experts

Scripts

This skill provides helper scripts in .agents/scripts/ for automating the MoE workflow:

Workflow

1. Define the Problem

Create a clear, specific prompt that all experts will analyze:

PROBLEM="Review the authentication flow in src/auth/ for issues and improvements"
TASK_ID=$(date +%s)

2. Spawn Experts in Parallel

Use moe-spawn.sh to spawn experts:

source .agents/scripts/moe-spawn.sh
spawn_moe_experts "architect security performance maintainer"

3. Wait for All Experts

Use moe-wait.sh to wait for completion:

source .agents/scripts/moe-wait.sh
wait_moe_experts "architect security performance maintainer"

4. Aggregate Results

Use moe-aggregate.sh to combine insights:

source .agents/scripts/moe-aggregate.sh
aggregate_moe_results "architect security performance maintainer"

5. Cleanup

Use moe-cleanup.sh to remove sessions and temp files:

source .agents/scripts/moe-cleanup.sh
cleanup_moe "architect security performance maintainer"

Complete Example: Code Review

Use the moe-code-review.sh script for a complete review workflow:

# Run MoE code review on a specific file
./.agents/scripts/moe-code-review.sh src/auth/login.ts

The script will:

1. Read the file content
2. Spawn architect, security, performance, and maintainer experts
3. Wait for all experts to complete
4. Synthesize results into actionable recommendations
5. Clean up sessions and temp files

Advanced: Weighted Aggregation

For weighted aggregation, modify the aggregator prompt in moe-aggregate.sh:

# Define weights
architect=3
security=3
performance=2
maintainer=2

# Build weighted prompt
WEIGHTED_PROMPT="Synthesize with these expert weights:\n"
for expert in $EXPERTS; do
  weight=$(eval echo \$$expert)
  WEIGHTED_PROMPT="$WEIGHTED_PROMPT\n- $expert (weight: $weight/10)"
done
WEIGHTED_PROMPT="$WEIGHTED_PROMPT\n\nHigher weight = more influence on final recommendation."

Best Practices

• Choose 3-5 experts — Too few misses angles, too many adds noise
• Make prompts specific — Generic prompts yield generic answers
• Include file content — Don't make experts hunt for context
• Define clear aggregation strategy — Consensus-based, weighted, or hierarchical
• Cache expert outputs — Save to files for inspection if aggregation fails
• Set timeouts — Kill hung experts after 5 minutes

Common Patterns

Limitations

• Expert outputs may conflict (aggregator must resolve)
• Token cost scales linearly with expert count
• No cross-expert communication during analysis
• Synthesis quality depends on aggregator prompt quality