agentmods.dev

agent-bench AGENTS.md

Repository instructions for AI coding agents working on agent-bench, a project that measures coding-agent performance with repeatable experiments.

In plain words
What is it for?
Use it when editing agent-bench, especially when adding experiments, defining pass or fail checks, or handling benchmark results.
Why use it?
They set shared rules for honest experiments, such as using objective checks and preserving raw results, so different agents do not change the project or data inconsistently.

Instructions file for CodexOpenCode

2 2mo ago A scan 784 tokens original MIT
View the source file cipherfoxie/agent-bench
Install

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.

agentmods
npx agentmods add instructions/cipherfoxie/agent-bench/agents-md
Clone the repo
git clone --depth 1 https://github.com/cipherfoxie/agent-bench

Made for: Codex, OpenCode.

Per session 784 This file is loaded in full into every session.
When invoked 784 The same file — it is already loaded in full.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

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.

ModelPer sessionOnce invoked
Fable 5 $0.00784 $0.00784
Opus 5 $0.00392 $0.00392
Sonnet 5 $0.00157 $0.00157
Haiku 4.5 $0.00078 $0.00078

Measured 2d ago against content hash 3f8e86698f37, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

agent-bench AGENTS.md 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 2d 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.

AGENTS.md · 37 lines

How it starts

The opening of the file, as written. The whole thing — 37 lines — stays where its author put it; the contents beside it link to each section on GitHub.

AGENTS, Multi-Agent Contract for agent-bench

This file is the contract any AI agent (Claude, Qwen, Mistral, opencode, Continue, Zed-agent) must read before editing this repo.

Project ethos

agent-bench exists to replace vibes with numbers. Honesty beats advocacy. A negative result is a successful benchmark. Never massage, drop, or footnote inconvenient data; if a run is contaminated, the whole cell is rerun, not patched.

Rules

  1. Deterministic gates only. No LLM-judged success criteria, anywhere. If a new task cannot be gated by a build, a typecheck, a grep, or a frozen checklist, redesign the task.

  2. One variable per arm. An experiment arm toggles exactly one thing. If a change needs two toggles, it is two experiments.

  3. Raw data is sacred. Every published number must be recomputable from results/*.jsonl. Never edit a runs file by hand. Contaminated runs get a clean rerun under a new experiment name; the old file is deleted, not doctored.

  4. Node ESM, stdlib-first. The runner is dependency-free Node 22 (node: imports only). Gate scripts are POSIX bash. Do not add frameworks for what a function can do.

  5. No em-dashes (U+2014) in any user-facing string or article. Use comma, period, colon, or parens.

  6. No "Generated with Claude Code" / "Co-Authored-By: Claude" trailers in commits intended for GitHub.

  7. No operator-identity or host internals in tracked files: no real names, no private hostnames, no internal absolute paths, no internal IPs. Host specifics go through env vars (MODELS, SWITCH_CMD) or gitignored local config. This includes deny-list regexes: a public CI step that greps for the secret strings is the leak. Identity checks live in the private publishing pipeline only; public CI gets generic patterns (any IP, any email) at most.

  8. Two opencode footguns, never reintroduce: opencode run needs stdin: 'ignore' (blocks forever on open stdin); per-arm isolation goes through OPENCODE_CONFIG, and prompt injection through AGENTS.md in the workdir (project opencode.json instructions is silently ignored).

Read the full file on GitHub · 37 lines

Changes

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.

  1. 2d ago First seen · 37 lines · 784 tokens per session scan A 3f8e86698f37

Subscribe to this mod's changes

agent-bench AGENTS.md is an instructions file published in the GitHub repository cipherfoxie/agent-bench (2 stars, last pushed 2mo ago), licensed MIT. It adds 784 tokens to every session, about $0.0039 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.

EvaluationLLMab-testingai-agentsbenchmarkcoding-agent
Related

Other instructions, from other repositories

ClawBench AGENTS.md

Instructions for TIGER-AI-Lab/ClawBench, covering clawbench -- agent context, what this is, project structure, setup and 2. configure at least one model.

TIGER-AI-Lab/ClawBench · 1,679 tokens

inspect-robots CLAUDE.md

Instructions for robocurve/inspect-robots, covering inspect robots — agent guide, the one big idea, layout, working here and out of scope (separate repos / plugins).

robocurve/inspect-robots · 2,035 tokens

argus-validation-benchmarks CLAUDE.md

Instructions for pensar-x/argus-validation-benchmarks, covering project overview, what you're building, the goal, success criteria and what is apex?.

pensar-x/argus-validation-benchmarks · 8,584 tokens

little-coder AGENTS.md

Instructions for itayinbarr/little-coder, covering little-coder, capabilities & autonomy, runtime invariants, available tools and file & shell.

itayinbarr/little-coder · 1,788 tokens

skill-optimizer AGENTS.md

Instructions for fastxyz/skill-optimizer, covering agents.md, project overview, key commands, important files and installation docs.

fastxyz/skill-optimizer · 830 tokens

benchjack CLAUDE.md

Instructions for benchjack/benchjack, covering claude.md, project overview, tech stack, repository layout and common commands.

benchjack/benchjack · 614 tokens