ai-agent-bench

ai-agent-bench is a skill for Claude Code from reidemeister94/development-skills. It costs 33 tokens per session (412 once invoked), scanned A, original, MIT.

A test harness for comparing Claude Code and Codex on the same real code-change task. It gives both agents the same starting code and checks whether their changes produce the required result.

In plain words
What is it for?
Use it to run sequential trials in separate worktrees, apply fast and full checks, preserve result branches, and produce a report comparing the agents.
Why use it?
It makes comparisons fairer by isolating each trial and recording outcome, elapsed time, token use, and cost. Failed trials are kept separate from successful comparisons.

Skill for Claude Code

Written for Claude Code: allowed-tools in frontmatter. Also seen: mentions Claude Code; mentions Codex.

Part of the development-skills plugin — 16 skills, 1 agent, 2 hooks shipped together

Good fit Use it to run sequential trials in separate worktrees, apply fast and full checks, preserve result branches, and produce a report comparing the agents.

Compare 6 skills from other repositories ↓
Install with agentmods
npx agentmods add skills/reidemeister94/development-skills/ai-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.

Any agent
npx skills add reidemeister94/development-skills --skill ai-agent-bench
Clone the repo
git clone --depth 1 https://github.com/reidemeister94/development-skills

Made for: Claude Code.

Or install development-skills, the plugin that ships this one along with the rest of its 16 skills, 1 agent, 2 hooks.

Wrote 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.

agentmods badge for ai-agent-bench

README.md
[![agentmods](https://agentmods.dev/badge/skills/reidemeister94/development-skills/ai-agent-bench/github.svg)](https://agentmods.dev/skills/reidemeister94/development-skills/ai-agent-bench)
Your own site
<a href="https://agentmods.dev/skills/reidemeister94/development-skills/ai-agent-bench"><img src="https://agentmods.dev/badge/skills/reidemeister94/development-skills/ai-agent-bench/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.

agentmods 80×15 button for ai-agent-bench

Your own site · 80×15
<a href="https://agentmods.dev/skills/reidemeister94/development-skills/ai-agent-bench"><img src="https://agentmods.dev/badge/skills/reidemeister94/development-skills/ai-agent-bench.svg" alt="Reviewed on agentmods" width="80" height="20"></a>
Per session 33 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 412 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. A grade says what 26 rules found in the file — not that it is safe.
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 5.1 $0.00033 $0.00412
Opus 5 $0.00016 $0.00206
Sonnet 5 $0.00007 $0.00082
Haiku 4.5 $0.00003 $0.00041

Measured 10d ago against content hash 10fb995f07f0, method: parsed. Prices are Anthropic first-party input rates as of 2026-09-10, from the pricing page.

Security

Grade A, and why

ai-agent-bench 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 10d ago.

The scan reads SKILL.md. This mod also ships 4 executable files (scripts/monitor.py, scripts/parse_transcript.py, scripts/run_evals.py, …), listed below but not scanned — reading those needs a real analyzer, not pattern matching.

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.

skills/ai-agent-bench/SKILL.md · 37 lines

What it actually says

AI agent bench

Compare agents only with the same task, starting commit, and outcome check. The harness preserves result branches and removes temporary worktrees.

Create <repo>/.agent-bench.toml:

prompt = "prompts/task.md"
start_branch = "main"              # or start_commit
agents = ["claude", "codex"]
outer_check = "./scripts/full_check.sh"
inner_check = "pytest tests/integration/test_x.py -q"

outer_check proves the real outcome before and after, and measures wall time. inner_check gives agents fast feedback.

Require a clean repo, available CLIs, and a passing outer_check. Confirm agents and run ID, then run trials sequentially to avoid load-biased timing:

python <skill>/scripts/run_trial.py --repo "$REPO" --config "$REPO/.agent-bench.toml" --agent "$AGENT" --run "$RUN_ID"

Results go to eval-results/<task>/<agent>/run-<id>-<timestamp>/. Record unexpected behavior in ai-agent-bench-anomalies.md per anomalies.

Aggregate with scripts/parse_transcript.py --aggregate <run-dirs> --output comparison.json --render-report comparison.md. Report gates, branches, time delta, tokens, and cost. Never rank a failed trial.

For plugin behavior rather than a real code task, use the bounded Pydantic runner documented by eval-regression and scripts/run_evals.py.

Never commit on the user's branch. A repeated run creates a new timestamped result and preserves prior evidence.

Files

What ships with it

6 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.

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. 10d ago First seen · 37 lines · 33 tokens per session scan A 10fb995f07f0

Subscribe to this mod's changes

ai-agent-bench is a skill published in the GitHub repository reidemeister94/development-skills (11 stars, last pushed 1mo ago), licensed MIT. It adds 33 tokens to every session and 412 once invoked, about $0.0002 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.

Related

Other skills, from other repositories

mutation-test

Mutation testing with two engines. Uses the project's NATIVE mutation runner (StrykerJS / Infection / mutmut / PIT / cargo-mutants) when one is configured — installing it on explicit consent when it is not — for a reproducible, comparable score; and an LLM-guided engine for the mutation classes native mutators cannot…

greglas75/zuvo · 190 tokens

write-tests

Write tests for existing production code. Processes ONE file at a time through a full pipeline: analyze, inventory (frozen BEFORE writing), write, executable coverage gate, verify, blind coverage audit, adversarial review, log. Uses CodeSift for discovery and analysis when available. Modes: [path] (specific target)…

greglas75/zuvo · 87 tokens

code-audit

Batch audit of production files against CQ1-CQ40 quality gates and CAP1-CAP29 anti-patterns. Tiered output (A/B/C/D), critical gate enforcement, evidence-backed scoring, cross-file pattern analysis, and prioritized execution plan. Flags: zuvo:code-audit all | [path] | [file] | --deep | --quick | --services |…

greglas75/zuvo · 83 tokens

plan

Analyzes architecture, selects patterns, assesses testability, then decomposes work into ordered TDD tasks with exact verification commands and explicit acceptance mapping. Works from an approved spec (zuvo:brainstorm output) or directly from a user-provided description.

greglas75/zuvo · 53 tokens

pentest

Hybrid white-box + black-box penetration testing across 7 dimensions (PT1-PT7). Stack-aware source-to-sink tracing, exploit verification, CMS overlay, and deterministic finding aggregation. Uses explicit candidate schemas, canonical-key deduplication, score caps, and MUST-GATE enforcement. Flags: zuvo:pentest [path] |…

greglas75/zuvo · 124 tokens

api-audit

API and endpoint integrity audit across 10 dimensions (D1-D10) plus optional contract stability (D11) and optional OWASP API Security Top 10 (D12: BOLA/BOPLA/BFLA, mass assignment, JWT alg-confusion, GraphQL introspection). Covers validation, payloads, pagination, errors, caching, HTTP semantics, waterfalls, rate…

greglas75/zuvo · 135 tokens