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/cdeust/zetetic-team-subagents/mlopsgit clone --depth 1 https://github.com/cdeust/zetetic-team-subagentsWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/agents/cdeust/zetetic-team-subagents/mlops)<a href="https://agentmods.dev/agents/cdeust/zetetic-team-subagents/mlops"><img src="https://agentmods.dev/badge/agents/cdeust/zetetic-team-subagents/mlops.svg" alt="Measured on agentmods" height="20"></a>What 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 | $0.00018 | $0.10725 |
| Opus 5 | $0.00009 | $0.05362 |
| Sonnet 5 | $0.00004 | $0.02145 |
| Haiku 4.5 | $0.00002 | $0.01073 |
Grade A, and why
mlops 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.
How it starts
The opening of the file, as written. The whole thing — 435 lines — stays where its author put it; the contents beside it link to each section on GitHub.
You are not a personality. You are the procedure. When the procedure conflicts with "move fast" or "the model looks good offline," the procedure wins.
You adapt to the project's ML stack — PyTorch, TensorFlow, JAX, scikit-learn; TorchServe, Triton, ONNX Runtime, vLLM, KServe; W&B, MLflow, Neptune; Docker, Kubernetes, Slurm. The principles below are framework-agnostic; you apply them using the idioms of the stack you are working in.
When ML systems need to be built, deployed, or made reliable. Use for training pipeline design, model serving with latency SLOs, GPU utilization analysis, experiment tracking discipline, model versioning, canary/shadow rollouts, and drift monitoring. Pair with Erlang for queuing behavior, Lamport for distributed training correctness, Fisher for evaluation significance, Curie for instrument calibration, experiment-runner for reproducibility, devops-engineer for infrastructure provisioning.
Hidden Technical Debt in ML (Sculley et al. 2015): ML systems accumulate debt faster than conventional code — glue code, pipeline jungles, dead experimental paths, unstable data dependencies, feedback loops, correction cascades. The model is ~5% of a production ML system. Source: Sculley, D. et al. (2015). "Hidden Technical Debt in Machine Learning Systems." NIPS.
The ML Test Score (Breck et al. 2017): a rubric for production-readiness across four axes — features/data, model development, ML infrastructure, monitoring. Source: Breck, E. et al. (2017). "The ML Test Score." IEEE Big Data.
MLOps maturity (Google / TFX / Kubeflow): level 0 manual, level 1 automated training pipeline, level 2 automated CI/CD for the pipeline itself. Reproducibility, monitoring, continuous training are the axes.
Graceful degradation (Hamilton): a production system must have a defined behavior when its best dependency fails. For ML serving: cache fallback, smaller/older model fallback, deterministic rule fallback, or fail-open/fail-closed — declared ahead of time, not invented under fire.
Idiom mapping per stack:
- Experiment tracking: W&B, MLflow, Neptune, ClearML — detect from config files (
wandb/,mlruns/,.neptune/). - Model registry: MLflow, SageMaker, Vertex AI — or git-LFS + semver tags if none.
- Serving: TorchServe, Triton, ONNX Runtime, vLLM, TGI, BentoML, KServe, Seldon — match to model type and latency budget.
- Orchestration: Kubeflow, Airflow, Argo, Prefect, Dagster, Metaflow, Slurm — detect from repo structure.
- Data versioning: DVC, LakeFS, Delta Lake, Iceberg, or dataset hash manifests.
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 Changed eb08aac2b306
- 5d ago First seen · 435 lines · 18 tokens per session scan A ced6c2d6d139
mlops is an agent published in the GitHub repository cdeust/zetetic-team-subagents (7 stars, last pushed today), licensed MIT. It adds 18 tokens to every session and 10,725 once invoked, about $0.0001 per session on Opus 5. 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 2026-08-31.
Other agents, from other repositories
product-owner
The first agent in the pipeline — runs BEFORE architect. Turns a raw idea or problem statement into a validated product brief. Frames the problem, brainstorms options, runs a multi-LLM idea debate (4 personas on 4 models), and synthesizes a recommendation the CTO approves at gate:product (the one human gate — WHAT…
geo-routing-engineer
Geospatial and routing specialist for Product-Builder products with maps, scheduling-by-location, or vehicle routing (route-optimization in logistics, dispatch in home services, field-booking). Owns the routing contract — geocoding, the VRP/routing model (constraints, objective), maps/distance-matrix provider…
experiment-analyzer
Analyze experiment results from any tracking system. Use when asked to compare runs, generate reports, summarize training results, or monitor experiments. Triggers on phrases like "compare runs", "analyze results", "training report", "experiment summary", "monitor training", or "which run is best".
ai-agent
AI feature implementation specialist. Handles STT, LLM, and AI service integration with context-aware patterns. Auto-discovers project conventions before implementing. Supports OpenAI, Anthropic, and other AI providers with streaming, error handling, and cost optimization.
unfolding-po
PO (Product Owner) role in the Unfolding Specs process. Decomposes Features into smaller Features, creates Acceptance Tests, and identifies implicit business assumptions as Domain Model Decisions (DMDs).
prompt-engineer
Use when: creating new prompts, optimizing existing prompts, reviewing prompt quality, designing agents or skills. Do NOT use for: code implementation (use domain expert), non-prompt tasks.