v3-performance-engineer — Claude Code subagent: performance targets
v3-performance-engineer — a subagent that benchmarks claude-flow v3 and pushes it toward stated goals: Flash Attention, HNSW search, cold start, memory. Use it when the numbers are claimed and need proving.
no license ★ 850 updated 2026-10-26
.claude/agents/v3-performance-engineer.mdWhat v3-performance-engineer does
Ships with ruvnet/agentic-flow. The job is to hit hard numbers on claude-flow v3: Flash Attention speedup in the 2.49x–7.47x band, AgentDB HNSW search 150x–12,500x faster than the linear pass, memory down by 50–75%, cold start under 500ms, SONA adaptation under 0.05ms.
Four benchmark groups come along. Cold start times the CLI, the MCP server and agent spawn. Search runs 10000 queries through both the old linear lookup and HNSW, then divides one by the other. Memory reads the heap before and after compression. Swarm coordination spawns 15 agents and times task decomposition plus consensus. Attention is measured on 512, 1024, 2048 and 4096 token sequences.
Both hooks are plain shell. The pre-hook prints the target list and checks whether npm and node exist on the machine. The post-hook saves a pattern with npx agentic-flow@alpha memory store-pattern, and a failure there is swallowed by || true — the run still finishes when npm is unreachable, you just lose the stored pattern.
What it helps with
- Benchmark cold start for the CLI, MCP server and agent spawn
- Compare HNSW search against linear lookup on 10000 queries
- Measure heap use before and after compression
- Time 15-agent swarm coordination and consensus
- Validate Flash Attention on long sequences
Example prompts for the agent
- Run v3-performance-engineer and benchmark cold start for the CLI and the MCP server.
- Compare linear search with HNSW on 10000 queries and show the speedup.
- Check Flash Attention on 512, 1024, 2048 and 4096 token sequences and store the result.
How to install v3-performance-engineer
Manually: put the files in place
.claude/agents/v3-performance-engineer.mdFor all projects use ~/.claude/agents/. Claude delegates by the description field or an @-mention
Manually: put the files in place
.codex/agents/v3-performance-engineer.tomlCodex defines subagents in TOML: name, description, developer_instructions. Move the Markdown text into developer_instructions by hand. For all projects use ~/.codex/agents/
Manually: put the files in place
.cursor/agents/v3-performance-engineer.mdCursor also reads .claude/agents/
Step by step, with commands for each platform: Claude Code subagents: how to create one, call it and keep context clean
Questions about v3-performance-engineer
How do I install v3-performance-engineer in Claude Code?
Take the agent file from ruvnet/agentic-flow and place it in your project at .claude/agents/v3-performance-engineer.md.
What does v3-performance-engineer do?
It benchmarks claude-flow v3 and checks the results against a target matrix covering Flash Attention, HNSW search, memory, cold start and swarm coordination.
What does the subagent need to run?
The pre-hook looks for npm and node; the post-hook calls npx agentic-flow@alpha memory store-pattern. Without reachable npm the pattern is not saved, but the run itself completes.