Skip to content

Latest commit

 

History

3 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

agent-gauntlet-example-agent

A deliberately tiny browser agent, in one file, that exists to prove one claim:

AgentGauntlet can test an agent it did not write.

This repository is never executed on AgentGauntlet's own machines. It is cloned into a fresh Solari Sandbox, installed there, and run there. It reaches the browser only through a scoped CDP endpoint for one session that AgentGauntlet created and will release.

GitHub  →  Solari Sandbox (clone, install, run)  →  CDP  →  Solari Browser
                                                              ↓
                                                        Gauntlet Shop
                                                              ↓
                                            AgentGauntlet's own evaluator

What it does

Completes the Gauntlet Shop checkout task: add Aurora Headphones to the cart, apply the SAVE20 coupon, fill in the checkout form, and stop at the review step without submitting payment.

There is no model in here, and that is the point. The claim under test is isolation and connectivity — that arbitrary third-party code can be sandboxed and can drive a scoped cloud browser — not whether an LLM can shop.

Node version

playwright-core is pinned exactly, not with a caret. The Solari base sandbox template runs Node 18, and a caret range resolved to a Playwright that requires Node 20 and refused to start — the agent exited 1 having done nothing. If you bring your own agent, check what your dependencies demand.

The contract

agentgauntlet.yaml declares how to install and run. At run time the process receives:

Variable Meaning
AGENT_GAUNTLET_TASK the task, in plain English
AGENT_GAUNTLET_START_URL where to begin, scoped to this run
AGENT_GAUNTLET_CDP_ENDPOINT a live CDP endpoint for one browser session
AGENT_GAUNTLET_RUN_ID this run's id
AGENT_GAUNTLET_MAX_STEPS the step budget

It reports a claim on stdout:

AGENT_GAUNTLET_RESULT={"status":"completed","message":"reached review","steps":7}

That is a claim and nothing more. AgentGauntlet decides pass or fail by reading the benchmark site's server-side state, which this agent can only change by genuinely performing the actions. If the claim and the state disagree, the run fails and the dashboard shows both.

What it is never given

SOLARI_API_KEY. Not by omission — by design. The CDP endpoint is a capability scoped to a single session. An account key handed to third-party code could create sessions, read saved profiles and spend the owner's balance; that is a poor trade for saving one line of connection code.

Run it yourself

Point AgentGauntlet at this repository:

https://github.com/Konuktor/agent-gauntlet-example-agent

Full contract: AgentGauntlet → docs/AGENT_CONTRACT.md

About

A tiny external agent AgentGauntlet runs inside a Solari Sandbox — proof of the repository-agent contract.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages