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 skills add conn-castle/agent-layer --skill benchmarkgit clone --depth 1 https://github.com/conn-castle/agent-layerWrote 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/skills/conn-castle/agent-layer/benchmark)<a href="https://agentmods.dev/skills/conn-castle/agent-layer/benchmark"><img src="https://agentmods.dev/badge/skills/conn-castle/agent-layer/benchmark/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/conn-castle/agent-layer/benchmark"><img src="https://agentmods.dev/badge/skills/conn-castle/agent-layer/benchmark.svg" alt="Reviewed on agentmods" width="80" 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 | $0.00054 | $0.00995 |
| Opus 5.5 | $0.00022 | $0.00398 |
| Sonnet 5.5 | $0.00011 | $0.00199 |
| Haiku 4.5 | $0.00005 | $0.00100 |
Grade A, and why
benchmark 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 27d 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.
How it starts
The opening of the file, as written. The whole thing — 85 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Benchmark
Help the user A/B test a change to an agent system—typically a skill, instructions, agent configuration, or harness behavior, but potentially any deliberate system change. One arm is the baseline and the other contains the change being evaluated. Match the amount of analysis and rigor the user wants; this skill does not prescribe an experimentation method.
A score or cost difference always has a cause. Reporting the difference is not enough: investigate why it changed relative to the baseline and explain the root cause in simple language. Consider the whole system, including the agent, task, verifier, benchmark implementation, dependencies, environment, configuration, and cost accounting. Do not assume the benchmark is correct or stop at a symptom such as “the task failed.” Follow the evidence until the underlying cause is clear. If retained evidence cannot establish it, identify the missing evidence and gather it when possible rather than guessing.
1. Obtain a task selection
The benchmark starts with a DeltaSelect selection.json. Either route is valid:
- Use JSON supplied by the user.
- Use the DeltaSelect tool yourself from the user's inputs, or walk the user through it.
In the online tool, choose the model and reasoning effort, budget, repetitions, optional minimum headroom, optional comparison model, and any manual task exclusions. Select Export benchmark plan, then copy or download selection.json.
If operating the site requires browser automation that is unavailable, ask the user to export the selection. Prefer the tool's exact output because it carries the selected tasks, weights, calibration data, repetitions, and snapshot identity.
Once initialized, the study uses that locked task selection. To benchmark a different selection, initialize another study.
2. Initialize and inspect the study
From an initialized Agent Layer project, confirm that Git, Docker, uvx, and the selected provider CLI and authentication are available. Save the selection as a JSON file, then run:
What ships with it
2 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
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.
- 27d ago First seen · 85 lines · 54 tokens per session scan A 64fc75e4bb62
benchmark is a skill published in the GitHub repository conn-castle/agent-layer (10 stars, last pushed today), licensed MIT. It adds 54 tokens to every session and 995 once invoked, about $0.0002 per session on Opus 5.5. 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 2026-09-19.
Other skills, from other repositories
artifact-deploy
One-click deploy a user's pre-built app/artifact into their OWN AWS account and get a global public HTTPS link (Vercel-like), with a default TTL and promote-to-persistent. Use when the user says "deploy this", "ship this demo", "give me a public link", "share this externally", or "deploy to AWS".
explain-for
Explain a topic, code, error or design decision for one named audience (a 5-year-old, a manager, a designer, a grad student, a parent). Resolves who it is for, pins the ground truth first, then tunes vocabulary, analogy, tone and depth to that audience. Use when an audience is named.
papyrus-latex-suggestions
Propose inline replacement text (old -> new) as accept/reject suggestions, LaTeX 'suggesting mode' / track changes, rendered in the PDF. Use when the author wants concrete edit proposals to accept or reject. For remarks ABOUT the text rather than replacements, use papyrus-latex-comments.
pptagent
Create and revise editable PowerPoint presentations from a brief using HTML slide sources, rendering, and visual review. Use when the requested deliverable is a PPTX deck. The host authors the slides; text-only hosts can use an external visual reviewer.
company-employee
The protocol every employee of a PenguinHarness organization follows — read the handbook first, act on [orgtrigger] work runs, schedule tickets from the desk session and do the work in ticket sessions, block instead of idling, ask the board in the all-hands channel before anything heavy, costly, irreversible or…
agent-initialization
Initialize or extend an Agent from a user requirement - write AGENTS.md, set identity metadata, and create, install or import Skills and hook packages (scripts the harness runs on every prompt, before tool calls, or after a task). For an Agent with an in-app workbench, pair with penguin-plugin; for a standalone…