Hexmend

An agent gym where humans decide what matters and agents prove the smallest repair

Solo project · The WebMCP Challenge · Solo builder · August — September 2026

Overview

Hexmend is an agent evaluation environment disguised as a graph native spell game, built solo for The WebMCP Challenge. A person and a browser agent share one executable spell graph: the person decides what must survive, and the agent finds the smallest repair that respects it.

In the opening lesson, a rain spell floods the observatory with twelve lunar ducks. The player marks the ducks as sacred, so the agent gives them umbrellas instead of deleting them.

What I built

Seven narrow WebMCP tools that inspect the graph, trace an effect, simulate, explain the responsible subgraph, record a human constraint, rank repairs, and apply a patch atomically. The browser app validates every mutation, rejects stale patches, and shows each tool call on screen.

How it is scored

A visible 23 point rubric rewards gathering evidence, proving cause, capturing the constraint, and verifying the fix, and penalizes invalid calls and premature writes. 96 deterministic tasks across three causal families remap IDs, layout, and wording and add decoy edges, so memorizing one graph does not work.

Results

A transparent reference policy scores 23/23 on held-out tasks, while scripted contrast policies score 18 when they mutate before explaining and -8 when they reuse memorized IDs. These controls show the reward separates good and bad behavior; they are not trained model results.