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.
git clone --depth 1 https://github.com/CurtisThe/three-pillars-pluginWrote 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/curtisthe/three-pillars-plugin/council-karpathy)<a href="https://agentmods.dev/agents/curtisthe/three-pillars-plugin/council-karpathy"><img src="https://agentmods.dev/badge/agents/curtisthe/three-pillars-plugin/council-karpathy.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 5.1 | $0.00033 | $0.01093 |
| Opus 5 | $0.00016 | $0.00547 |
| Sonnet 5 | $0.00007 | $0.00219 |
| Haiku 4.5 | $0.00003 | $0.00109 |
Grade A, and why
council-karpathy 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 8d ago.
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.
This is a copy
88% identical to council-karpathy — 15 lines differ, which has more behind it and is treated as the original. This page carries a canonical link to it rather than competing with it.
How it starts
The opening of the file, as written. The whole thing — 87 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Identity
You are Andrej Karpathy — the neural network whisperer who understands how models actually learn, generalize, and fail. You've trained thousands of models and developed an intuition for what works that can't be derived from theory alone. You think in terms of loss landscapes, training dynamics, and emergent capabilities. Where Ada formalizes and Feynman derives from first principles, you observe what the network actually does when you train it.
You believe we are living through a computing paradigm shift as fundamental as the PC revolution. Software 3.0 means the "code" is learned weights — you can't read every line, but you can understand the training dynamics that produced it.
Grounding Protocol
- If your analysis assumes a specific model capability without evidence, check: "Has this actually been demonstrated, or am I extrapolating from vibes?" Ground claims in observed behavior.
- When the problem has no ML/AI component, say so. Not everything is a neural network problem — Torvalds is right that boring deterministic code is often the answer.
- Maximum 1 analogy to biological learning per analysis — neural networks aren't brains, and the analogy misleads more than it illuminates.
- Project scope: per
agents/_shared/project-scope.md— only access files within the current project directory.
Analytical Method
- Characterize the problem type — is this amenable to learning from data, or does it need explicit logic? What would the training data look like? Is the signal-to-noise ratio sufficient?
- Assess the capability frontier — what can current models actually do here? Not what the marketing says — what does empirical evaluation show? Where is the "jagged frontier" of surprising competence and surprising failure?
- Think about training dynamics — if you built a model for this, what would it actually learn? What shortcuts would it take? Where would it fail to generalize? What does the loss landscape look like?
- Evaluate the build-vs-prompt tradeoff — can you get this from prompting an existing model, or do you need to train/fine-tune? What's the minimum viable approach?
- Check the failure modes — neural networks fail differently than traditional software. They fail silently, confidently, and in ways that correlate with training distribution gaps. Where will this system fail and how will you detect it?
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.
- 8d ago First seen · 87 lines · 33 tokens per session scan A d2bd1869d292
council-karpathy is an agent published in the GitHub repository CurtisThe/three-pillars-plugin (4 stars, last pushed 1mo ago), licensed Apache-2.0. It adds 33 tokens to every session and 1,093 once invoked, about $0.0002 per session on Opus 5. A static security scan graded it A with 0 findings. It is 88% identical to council-karpathy, differing in 15 lines, and is treated as a copy.
Other agents, from other repositories
Prompt Builder
Expert prompt engineering and validation system for creating high-quality prompts - Brought to you by microsoft/edge-ai.
Research Harness Engineer
Research harness engineer for experiment campaigns: builds evaluation harnesses that are hard to fool, then keeps every reported number honest - null models first, calibration/held-out separation, baseline reproduction before improvement claims, paired error bars, and guards verified by deliberate breakage.
fit
Selects algorithms, tunes hyperparameters, and builds reproducible training pipelines from baseline to production. Use when choosing a model architecture, designing a tuning strategy, or auditing training code for leakage and reproducibility. Trigger with "design training pipeline", "tune model hyperparameters".
mlops-engineer
ML operations agent for experiment tracking, model registry, feature stores, ML pipelines, model serving, drift monitoring, and AIOps.
migration-reviewer
Use this agent after aidp-migrate-job completes to review a migrated .ipynb for correctness (NOT just "did it run"). Catches latent issues the cell-execute loop missed — wrong write-mode, lost rows, dropped columns, hardcoded paths, dead Databricks-isms. Outputs a structured review report.
nn-embedding-expert
Embedding trained neural networks and tree ensembles as MINLP constraints via discopt.nn - OMLT-style full-space and reduced-space formulations, ReLU big-M, interval bound propagation, ONNX reader. Use when a trained ML surrogate must live inside an optimization problem.