agent-benchmark

A self-benchmarking workflow in which an agent writes code and tests, receives adversarial reviews from multiple providers, and fixes the findings.

In plain words
What is it for?
Use it to benchmark code quality, review quality, and test quality, with options for quick runs, skipped tests, or a dry run.
Why use it?
It gives you a repeatable way to assess and improve coding-agent work across different models.

Skill for Claude CodeCodex

Part of the zuvo plugin — 34 skills, 21 agents, 5 hooks shipped together

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 skills/greglas75/zuvo/agent-benchmark
Any agent
npx skills add greglas75/zuvo --skill agent-benchmark
Clone the repo
git clone --depth 1 https://github.com/greglas75/zuvo

Made for: Claude Code, Codex.

Or install zuvo, the plugin that ships this one along with the rest of its 34 skills, 21 agents, 5 hooks.

Per session 61 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 3,093 The whole file, excluding the scripts and references it only reads on demand.
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 $0.00061 $0.03093
Opus 5 $0.00030 $0.01546
Sonnet 5 $0.00012 $0.00619
Haiku 4.5 $0.00006 $0.00309

Measured 3d ago against content hash c67e093f7c13, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

agent-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 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.

skills/agent-benchmark/SKILL.md · 341 lines

How it starts

The opening of the file, as written. The whole thing — 341 lines — stays where its author put it; the contents beside it link to each section on GitHub.

zuvo:agent-benchmark — Self-Benchmark

You are the subject of this benchmark. YOU write the code and tests. Adversarial review (multi-provider) critiques your work between rounds. You fix based on findings.

Run this skill in different models (Opus, Sonnet, Haiku) to compare agent quality.

Argument Parsing

Flag Effect
--quick Skip adversarial rounds (R1 code + R3 tests only, no fixes)
--no-tests Skip test rounds (R1 + R2 only)
--dry-run Print what would happen, don't execute
(no flags) Full 4-round benchmark with adversarial

Mandatory File Loading

Read these files before starting:

CORE FILES LOADED:
  1. ../../shared/includes/benchmark-corpus/task-code.md    -- READ/MISSING
  2. ../../shared/includes/benchmark-corpus/task-tests.md   -- READ/MISSING
  3. ../../shared/includes/benchmark-scoring-rubric.md      -- READ/MISSING
  4. ../../shared/includes/run-logger.md                    -- READ/MISSING
  5. ../../shared/includes/retrospective.md                    -- READ/MISSING

If any file is missing, stop.


Setup

  1. Detect current model: check $CLAUDE_MODEL or infer from context. Record as agent_model. Build a short slug using this table:

    Model string contains Slug
    opus opus
    sonnet sonnet
    haiku haiku
    gpt-5.4 or codex + 5.4 codex-5.4
    gpt-5.3 or codex + 5.3 codex-5.3
    gpt-5 or codex (no version) codex
    gemini gemini
    composer composer
    cursor cursor
    anything else use the model name as-is

    Do NOT use agent as a slug. The slug must identify which model ran the benchmark.

  2. Create output directory with agent name in folder:

    AGENT_SLUG="<slug from step 1>"
    RUN_ID="agent-bm-${AGENT_SLUG}-$(date -u +%Y%m%dT%H%M%SZ)"
    OUT_DIR="zuvo/reports/${RUN_ID}"
    mkdir -p "$OUT_DIR"
    

Dispatch is already authorized — do not ask, do not downgrade. Invoking this skill IS the request for every agent and gate it mandates, so a session rule about unprompted Agent use does not apply here. Only a harness with NO dispatch capability takes the documented single-agent fallback, and it still runs every gate inline — see ../../shared/includes/env-compat.md. Skipping a mandated agent and self-scoring the result is a substituted gate, not a degraded run.

Read the full file on GitHub · 341 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. 3d ago First seen · 341 lines · 61 tokens per session scan A c67e093f7c13

Subscribe to this mod's changes

agent-benchmark is a skill published in the GitHub repository greglas75/zuvo (6 stars, last pushed 3d ago), licensed MIT. It adds 61 tokens to every session and 3,093 once invoked, about $0.0003 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.

Related

Other skills, from other repositories

roam

Codebase comprehension via roam-code CLI. Use when exploring codebases, planning modifications, debugging failures, assessing PR risk, or checking architecture health. Triggers on: understanding project structure, pre-change safety checks, finding symbols/files, blast radius analysis, affected tests, health scoring…

Cranot/roam-code · 86 tokens

ring:writing-plans

Writing a rolling-wave phased implementation plan from a spec before coding: a phase-epic-task hierarchy where Phase 1 is detailed into dispatch-ready tasks and later phases stay epic-level for elaboration during execution. Use when a multi-file feature needs decomposition; runs after ring:exploring-codebases or…

LerianStudio/ring · 94 tokens

ring:searching-code

Forensic code search and analysis with optional Chain of Draft (CoD) ultra-concise mode. Five-phase methodology (clarification, planning, execution, analysis, synthesis) with severity assessment. Use for targeted investigation of specific patterns, bugs, or vulnerabilities. Skip for broad architecture mapping (use…

LerianStudio/ring · 74 tokens

ring:exploring-codebases

Exploring a codebase across phases: scopes the target, detects architecture, components, and layers, deep-dives each discovered perspective, then synthesizes findings into actionable guidance with file:line evidence. Use to understand how a feature or system works before planning changes, or to orient on an unfamiliar…

LerianStudio/ring · 91 tokens

ring:generating-release-guides

Generating an internal Operations-facing update/migration guide from the git diff between two refs, documenting per-change client impact, deploy ordering, monitoring, and rollback notes in English, pt-br, or both. Use when preparing a version release or recording what changed for the Ops team. Runs read-only by…

LerianStudio/ring · 85 tokens

ring:writing-skills

Writing or editing a Ring skill: SKILL.md structure, frontmatter and Agent-Search-Optimization rules, token-efficiency targets, and bulletproofing (Iron Law, rationalization tables, Red Flags) so discipline-enforcing skills resist excuses. Use when creating or revising a skill. Delegates pressure-testing to…

LerianStudio/ring · 100 tokens