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 instructions/minghinmatthewlam/openbench/agents-mdgit clone --depth 1 https://github.com/minghinmatthewlam/openbenchWhat 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 | $0.02496 | $0.02496 |
| Opus 5 | $0.01248 | $0.01248 |
| Sonnet 5 | $0.00499 | $0.00499 |
| Haiku 4.5 | $0.00250 | $0.00250 |
Grade A, and why
openbench 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 3d 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 — 181 lines — stays where its author put it; the contents beside it link to each section on GitHub.
OpenBench — Agent Context
Read this first. It captures what this project is trying to become, so any agent
or contributor picks up the strategic context, not just the mechanics in
README.md / WRITEUP.md.
Local execution context
Before changing benchmark execution or starting a run, read agents.env when
it exists. It is a gitignored, machine-local source of truth for where code is
developed and where benchmarks are executed. Never commit it or put credentials
in it.
For this installation, code changes belong in the laptop checkout. Benchmark runs normally execute on the Mac Mini from an exact pushed commit. Do not edit source on the Mini, do not launch from a dirty or stale checkout, and check for active benchmark processes before starting another run.
What OpenBench is
A benchmark framework for comparing coding-agent harnesses (codex, pi,
opencode, cursor, devin, claude, ...) — the CLI products that wrap a model in a
run loop, tool set, and permission policy. Tasks are self-contained
(task.toml + instruction.md + Docker environment + verifier); the verifier
is the sole judge.
Execution ownership
The canonical path is obench run [suite.toml]. OpenBench compiles immutable
suite intent, stock/custom profiles, one exact Harbor job per task set, and
comparison-plan sidecars. Pinned Harbor owns task/trial execution, Docker
sandboxes, concurrency, retries, resume, locks, verifier execution, and ATIF
artifacts. OpenBench then validates/imports every intended job as one atomic
suite result, reconciles optional proxy evidence, and owns comparison,
statistics, publication policy, and site acceptance.
obench legacy run is compatibility only. Manual Harbor export, job-run, and
result-import commands remain diagnostics and migration tools; do not describe
them as the default workflow.
Product goals (the two things we are building toward)
- Community harness flywheel. Make OpenBench trivially importable and usable so third parties can add their own harnesses or harness variations and evaluate them against the stock adapters. If someone builds a better harness or feature, they should want to use OpenBench to prove it and post the results publicly — that showing-off loop is how the framework grows.
- Company/private-codebase evals. More teams evaluate agents on their own codebases and use cases rather than general benchmarks. OpenBench should be easily installable inside a private repo so companies can benchmark harnesses and models on their tasks with the same rigor (checker polarity, token metering, Wilson CIs) as the public tiers.
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.
- 3d ago First seen · 181 lines · 2,496 tokens per session scan A dafb3b8654e9
openbench AGENTS.md is an instructions file published in the GitHub repository minghinmatthewlam/openbench (131 stars, last pushed 6d ago), licensed MIT. It adds 2,496 tokens to every session, about $0.0125 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 2026-08-30.
Other instructions, from other repositories
benchflow AGENTS.md
Instructions for benchflow-ai/benchflow, covering benchflow, setup + test, conventions and skill catalog (.agents/skills, mirrored at .claude/skills).
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?.
little-coder AGENTS.md
Instructions for itayinbarr/little-coder, covering little-coder, capabilities & autonomy, runtime invariants, available tools and file & shell.
tapp AGENTS.md
AGENTS.md instructions for aarwitz/tapp, covering tapp agent playbook, no mcp connected? just run the cli, pick the right tool for the job, session driving (the playwright loop) and autonomous exploration (tappexplore).
superpowers-evals CLAUDE.md
Claude Code instructions for prime-radiant-inc/superpowers-evals, covering superpowers evals, canonical actors, commands, architecture and scenario conventions.
superpowers-evals AGENTS.md
AGENTS.md instructions for prime-radiant-inc/superpowers-evals, a project described as: Behavioral eval lab (Quorum) for the superpowers project that drives real coding-agent CLIs (Claude, Codex, Gemini, Kimi, and more) through a QA agent and grades them on workflow compliance against scenario criteria and…