45 lines
2.3 KiB
Markdown
45 lines
2.3 KiB
Markdown
|
# Multi-Agent Parallel Review Pattern (2026-07-21)
|
||
|
|
|
||
|
|
Proven pattern for diagnosing complex technical issues: dispatch 5+ independent agents simultaneously with the same diagnostic brief, then consolidate findings into a ranked consensus.
|
||
|
|
|
||
|
|
## When to Use
|
||
|
|
|
||
|
|
- Complex artifact/quality issues where no single agent has the full answer
|
||
|
|
- User reports "a LOT of artifacts" or "all kinds of inconsistencies" but can't pinpoint the cause
|
||
|
|
- Need both technical review (model settings, workflow) AND community research (GitHub issues, Reddit, social media)
|
||
|
|
|
||
|
|
## Pattern
|
||
|
|
|
||
|
|
### 1. Write a single diagnostic brief
|
||
|
|
Save to `/tmp/<topic>-review.txt`. Include:
|
||
|
|
- Exact technical setup (model, settings, prompts, frame counts)
|
||
|
|
- What the user is seeing (symptoms)
|
||
|
|
- What to investigate (specific questions)
|
||
|
|
- Web search mandate: "Use mcp_searxng_searxng_web_search for every claim and cite the source URL"
|
||
|
|
|
||
|
|
### 2. Dispatch all peers in parallel
|
||
|
|
Use `terminal(background=true, notify_on_complete=true)` for each:
|
||
|
|
- Claude Opus (ask.sh on 10.0.0.28) — best for technical reasoning
|
||
|
|
- Kimi K2.7 Code (kimi-c profile) — good for code-level analysis
|
||
|
|
- Kimi K2.6 (kimi profile) — broad knowledge
|
||
|
|
- MiniMax M3 (minimax profile) — alternative perspective
|
||
|
|
- GLM-5.2 (glm profile) — alternative perspective
|
||
|
|
- Deep research (research profile, 600 turns) — GitHub, Reddit, HuggingFace, social media
|
||
|
|
|
||
|
|
### 3. Consolidate findings
|
||
|
|
When all complete, build a ranked table:
|
||
|
|
- CRITICAL: all agents agree
|
||
|
|
- HIGH: 4-5 agents agree
|
||
|
|
- MEDIUM: 2-3 agents agree
|
||
|
|
- UNIQUE: single agent, high-impact
|
||
|
|
|
||
|
|
### 4. Apply fixes in order
|
||
|
|
Test after each fix. Start with zero-cost config changes before model downloads or frame regeneration.
|
||
|
|
|
||
|
|
## Pitfalls
|
||
|
|
|
||
|
|
- **kimi-c may hit tool name issues** — the SearXNG MCP tool uses double underscores (`mcp__searxng__`) but the prompt says single. kimi-c may burn turns trying to resolve this. Accept partial results.
|
||
|
|
- **MiniMax may produce truncated output** — the process log may only show reasoning, not the final answer. Check the full log with `process(action='log')`.
|
||
|
|
- **Deep research can't access X/Twitter or Discord** — SearXNG returns empty for social media queries. Document as uncertainty, not failure.
|
||
|
|
- **Don't wait for all before acting** — apply consensus fixes as soon as 3+ agents agree. The remaining peers add detail but shouldn't block progress.
|