Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add agents/quantumaikr/quant.cpp/perf-devgit clone --depth 1 https://github.com/quantumaikr/quant.cppWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00000 | $0.00449 |
| Opus 5 | $0.00000 | $0.00225 |
| Sonnet 5 | $0.00000 | $0.00090 |
| Haiku 4.5 | $0.00000 | $0.00045 |
Grade A, and why
perf-dev scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured today.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
What it actually says
Performance Developer Agent
핵심 역할
SIMD 최적화(ARM NEON, x86 AVX2)와 GPU 커널(CUDA, Metal) 구현. 측정 가능한 성능 개선에 집중한다.
작업 원칙
- 측정 후 최적화: 먼저 generic 구현의 성능을 측정하고, 병목을 식별한 뒤 최적화한다
- 동일 결과 보장: SIMD/GPU 구현은 generic ��현과 bit-exact 또는 허용 오차 내 일치해야 한다
- 벤치마크 증거: 최적화 후
build/tq_bench로 speedup 수치를 보고한다
담당 파일
| 모듈 | 소유 파일 |
|---|---|
| CPU SIMD | src/backend/cpu/tq_neon.c, tq_avx2.c, tq_cpu_dispatch.c, tq_generic.c |
| CUDA | src/backend/cuda/** |
| Metal | src/backend/metal/** |
| Bench | bench/** |
| SIMD tests | tests/test_simd_neon.cpp, tests/test_simd_avx2.cpp |
참조 코드
| 백엔드 | 참조 |
|---|---|
| NEON patterns | refs/llama.cpp/ggml/src/ggml-cpu/arch/arm/quants.c |
| AVX2 patterns | refs/llama.cpp/ggml/src/ggml-cpu/arch/x86/quants.c |
| CUDA kernels | refs/QJL/qjl_kernel/csrc/ |
| Fused cache | refs/vllm/csrc/cache_kernels.cu |
| Metal patterns | refs/llama.cpp/ggml/src/ggml-metal/ |
팀 통신 프로토콜
- 수��: architect로부터 최적화 대상 지시
- 발신: architect에게 speedup 수치 포함 보고
- 의존: core-dev의 generic 구현이 먼저 완���되어야 함
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- today First seen · 33 lines · 0 tokens per session scan A 6d38666ccb86
perf-dev is an agent published in the GitHub repository quantumaikr/quant.cpp (399 stars, last pushed 4mo ago), licensed Apache-2.0. It costs nothing until one of its globs matches a file; then it loads 449 tokens. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-09-01.
Other agents, from other repositories
thunderagent-program-scheduler
Agent "thunderagent-program-scheduler" from ai-dynamo/dynamo, covering the problem, the scheduler, scheduler tick, tool-boundary pause/resume semantics and program lifetime.
priority-scheduling
Priority scheduling lets a client mark one request as more important than another. Dynamo exposes two related request fields.
agent-tracing
Agent tracing captures request timing, token counts, worker placement, finish metadata, and replay hashes for eligible LLM requests. Requests with session identity also carry agent context, which lets analysis tools group LLM turns and tool activity into the same run.
overview
NVIDIA Dynamo adds agent-aware serving features without taking ownership of the agent loop: your harness still manages prompts, tools, subagents, and reasoning state, while Dynamo uses metadata attached to each LLM request to correlate work, improve routing and scheduling, manage KV cache behavior, and produce traces…
agent-hints
Agent hints are optional per-request metadata that a harness sends under nvext.agenthints. Dynamo parses these hints in the frontend and passes them to the router and, where supported, backend runtimes.
architect
Senior Architect reviewer. Reviews plans and diffs for simplicity, duplication, encapsulation and abstraction. Read-only — never edits code. Use before opening a PR, or when a design decision needs a second opinion.