Skip to content

Latest commit

 

History

38 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 

Repository files navigation

agent-bench

A benchmark repository for comparing coding-agent configurations. One skill drives the protocol, a catalog of cases supplies the work, and no executable harness is maintained in this repository.

Install

npx skills add collectiveai-team/agent-bench

The CLI clones the repository and copies skills/agent-bench/ — including the full cases/ tree — into the skills directory for your harness. Where that goes depends on context:

  • Run from inside a project directory: installs to .agents/skills/agent-bench/ in that project and creates a symlink at .claude/skills/agent-bench.
  • Run from outside any project (global install): installs to ~/.agents/skills/agent-bench/ and symlinks into ~/.claude/skills/agent-bench.

In both cases SKILL.md, references/, templates/, and cases/ (all seven cases plus _base-taskflow) are present at the installed path.

Install caveats

npx skills add was verified against collectiveai-team/agent-bench on 2026-08-10, run from inside a project directory. The complete cases/ directory traveled with the install; no selective-copy behaviour was observed. All eight case directories (_base-taskflow, B-sabotage, L-hookrelay, L-taskflow, M-relay, R-envelope, S-ingest, S-ledger) were present at the installed path alongside SKILL.md, references/ and templates/.

Only the project-local install path was exercised. The global path is documented from the CLI's behaviour with other skills on the same machine, not from a run against this repository.

If you need the skill accessible without running npx skills add — for example, in a CI environment — clone the repository and point $AGENT_BENCH_HOME at it:

git clone git@github.com:collectiveai-team/agent-bench.git ~/agent-bench
export AGENT_BENCH_HOME=~/agent-bench

The run ledger resolves $AGENT_BENCH_HOME before falling back to local discovery (see ## Ledger location below).

Use

Invoke the agent-bench skill inside your harness and name the two arms to compare. The skill runs both arms against the same case and scores the results.

Example: "Compare orq-lite factory_governed with fast=true against fast=false on S-ledger, N=5."

What is in here

Path Contents
skills/agent-bench/ The protocol and the case catalog
runs/ Committed evidence, one directory per run
docs/superpowers/ Design spec and implementation plan

Ledger location

Runs are written to the first location that resolves:

  1. $AGENT_BENCH_HOME if set
  2. The current repository, if it is an agent-bench checkout
  3. ./bench-runs/warning: results written here are not recorded to the catalog

Design

docs/superpowers/specs/2026-08-07-agent-bench-design.md

About

No description, website, or topics provided.

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages