Who it is forA researcher testing a conjecture who wants an agent to propose and check candidate counterexamples.
The problemChecking many candidate counterexamples by hand is slow, and the checks need a dedicated environment separate from ordinary code work.
The outcomeA set of candidate counterexamples checked in the Counterexample Lab, with the Reviewer's verdict on each result.
What you need first
- Node.js 22.12 or later
- One authenticated Agent CLI, such as GitHub Copilot CLI, OpenAI Codex CLI, or Claude Code
- A project folder you are willing to let the agents modify
Steps
- Complete the basic install: Node.js 22.12 or later and one authenticated Agent CLI.
- Open the setup guide at docs/counterexample-lab-jacobian.md and follow it to start the Counterexample Lab and the isolated Jacobian MCP bridge.
- Run
argusand describe the conjecture and the search space to the Planner. - Let the Engineer run the candidate checks and review the Reviewer's verdicts.
- Use
argus --statusto confirm which candidates are verified and which are blocked.
Prompt for your agent
Using the microsoft/ArgusAgent repo, set up the Counterexample Lab from docs/counterexample-lab-jacobian.md and search for counterexamples to the conjecture <conjecture statement>. Expect a list of candidate counterexamples with a Reviewer verdict for each. Verify that every reported counterexample was checked by the lab, not only proposed by the model.
Caveat. The Jacobian bridge and lab setup live in a separate guide, so the steps here depend on that document, which this summary did not read in full.
Use this in your AI agent
Ask your agent
Using the Deduped MCP, get the setup steps for microsoft/ArgusAgent and "Search for mathematical counterexamples with the Counterexample Lab"
Add Deduped to Claude Code once
claude mcp add --transport http deduped https://dedupedai.com/mcp
Every Monday: the 5 hottest repos and exactly how to use them. Free, one email a week, unsubscribe anytime.
Get them freeShare: X LinkedIn Hacker News Email Copy link
Third-party repos we don't own. AI-written use cases can be wrong. Disclaimer