build-agent-evals

build-agent-evals is a skill for Claude Code, Codex from Hoja-Solutions/agent-stdlib. It costs 175 tokens per session (1,353 once invoked), scanned A, original, MIT.

A method for building automated tests that measure whether an AI agent improves or regresses after changes.

In plain words
What is it for?
Use it to collect failure cases, create reference solutions, choose graders, isolate test runs, compare pass rates, and expand the test suite over time.
Why use it?
It replaces guesses based on a few examples with repeatable checks based on real failures and clearly defined pass or fail decisions.

Skill for Claude CodeCodex

Part of the agent-stdlib plugin — 14 skills, 2 commands, 1 agent, 2 hooks, 2 MCP servers 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/hoja-solutions/agent-stdlib/build-agent-evals
Any agent
npx skills add Hoja-Solutions/agent-stdlib --skill build-agent-evals
Clone the repo
git clone --depth 1 https://github.com/Hoja-Solutions/agent-stdlib

Made for: Claude Code, Codex.

Or install agent-stdlib, the plugin that ships this one along with the rest of its 14 skills, 2 commands, 1 agent, 2 hooks, 2 MCP servers.

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 build-agent-evals

README.md
[![agentmods](https://agentmods.dev/badge/skills/hoja-solutions/agent-stdlib/build-agent-evals.svg)](https://agentmods.dev/skills/hoja-solutions/agent-stdlib/build-agent-evals)
Your own site
<a href="https://agentmods.dev/skills/hoja-solutions/agent-stdlib/build-agent-evals"><img src="https://agentmods.dev/badge/skills/hoja-solutions/agent-stdlib/build-agent-evals.svg" alt="Measured on agentmods" height="20"></a>
Per session 175 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 1,353 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 $0.00175 $0.01353
Opus 5 $0.00088 $0.00677
Sonnet 5 $0.00035 $0.00271
Haiku 4.5 $0.00017 $0.00135

Measured 4d ago against content hash cc95e48254a9, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

build-agent-evals 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 4d ago.

The scan reads SKILL.md. This mod also ships 1 executable file (scripts/passk.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/build-agent-evals/SKILL.md · 80 lines

How it starts

The opening of the file, as written. The whole thing — 80 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Build agent evals

Source: Demystifying evals for AI agents, with the section on keeping evals honest drawing on Eval awareness in BrowseComp and Designing AI-resistant technical evaluations. A standalone gist of the core material exists but is undiscoverable; this skill packages it and adds the runnable metric script.

An eval tells you whether a change to an agent made it better or worse. Without one you are guessing from vibes, and vibes miss regressions that only show up on the tenth run. Treat the eval suite the way you treat a unit-test suite: it has an owner, it grows when bugs slip through, and it fails loudly.

Start from real failures

Collect 20 to 50 tasks before writing any grader. The best sources are bugs your agent already produced, support tickets, and manual test cases you keep rerunning by hand. Write each task so two experts reading it reach the same verdict on pass or fail. If you cannot decide whether an output passed, the task is underspecified and will poison every measurement built on it.

Include a reference solution for each task to prove it is solvable, and build both positive cases (the agent should do X) and negative cases (the agent should refuse, or should not touch Y). A suite made only of positive cases optimizes toward an agent that does too much.

Choose the grader to match the task

Grade what the agent produced, not the path it took. An agent that reaches the right end state by an unusual route still passed.

  • Code-based grader. String match, schema validation, a state check against a database or filesystem. Use this wherever the correct answer is checkable by a program. It is fast, free, and never flaky in the way a model judge is.
  • Model-based grader (LLM-as-judge). A rubric scored by a separate model call. Use it for output that needs judgment: tone, summary quality, whether an explanation is correct. Give the judge a rubric with explicit criteria rather than asking "is this good," and have it cite evidence for its score so you can audit it.
  • Human grader. Subject-matter spot checks and A/B preference. Use it sparingly to calibrate the other two, not as the everyday loop.

Read the full file on GitHub · 80 lines

Files

What ships with it

1 file 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. 4d ago First seen · 80 lines · 175 tokens per session scan A cc95e48254a9

Subscribe to this mod's changes

build-agent-evals is a skill published in the GitHub repository Hoja-Solutions/agent-stdlib (1 stars, last pushed 1mo ago), licensed MIT. It adds 175 tokens to every session and 1,353 once invoked, about $0.0009 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

unified_meta_agent_skill

A foundational meta-agent skill for generating Multi-Agent Systems (MAS). It systematically drives the process from conceptual task decomposition to agent engineering, and finally to workflow orchestration.

linhh29/Skill_MAS · 39 tokens

controlflow-verify

Use after /controlflow-claude-code:controlflow-plan produces a plan, before implementation. Runs adversarial pre-execution verification inline in the main context (zero subagents): structural audit, assumption/mirage detection, and executability cold-start simulation. Emits APPROVED / NEEDSREVISION / REJECTED with…

Smithbox-ai/ControlFlow · 73 tokens

controlflow-planning

Use when a repository task needs a strict ControlFlow-style implementation plan before coding, especially for small, medium, or large scope, cross-file edits, risky migrations, architectural uncertainty, or any work that should produce a saved Markdown plan artifact in plans/.

Smithbox-ai/ControlFlow · 55 tokens

controlflow-router

Use when a task broadly matches ControlFlow for Codex and you need to decide whether to start with spec capture, strict workflow, strict planning, pre-execution plan review, orchestration, code review, or memory hygiene, or whether to combine several of those skills in sequence.

Smithbox-ai/ControlFlow · 60 tokens

controlflow-plan

Use when a repository task needs a high-quality ControlFlow plan before coding — small, medium, or large scope, cross-file edits, risky migrations, architectural uncertainty, or any work that should produce a saved Markdown plan artifact in plans/. Single-sources the plan format from schemas/planner.plan.schema.json…

Smithbox-ai/ControlFlow · 69 tokens

controlflow-spec

Use before planning non-trivial Codex work when the requirements, acceptance criteria, boundaries, constraints, or success measures are not yet clear enough for a safe implementation plan.

Smithbox-ai/ControlFlow · 38 tokens