A model on its own answers questions; a harness turns it into an agent that reads files, runs tools and finishes work. Intelligence is models × hardware × harnesses, and this is the third factor: who builds the agents, how fast they ship, how many people use them, and how much they add to the model inside.
Every harness placed by two freedoms: can you read and change its code (left to right), and can you choose the model inside it (bottom to top)? Bubble size follows the measure picked in the race below. Filters re-arrange the map; hover a harness for its licence, pricing and sources.
The same harnesses ranked seven ways: first across all measures, then one measure at a time. Switch the measure and watch them re-sort: popularity among developers (stars), real use (downloads, editor installs, tokens through OpenRouter) and how fast they ship are different races, with different leaders.
Weekly package downloads (npm or PyPI) on a log scale, where equal steps mean equal multiples. Hover a line, or a name in the race above, to follow one harness.
Downloads are a proxy for use, not users: continuous integration inflates them, and harnesses installed with a native installer (Claude Code today, Cursor, Grok Build, Goose, Warp) are partly or wholly invisible here. npm keeps 18 months of history, PyPI six; the site archives every week from now on.
Each dot is one harness running one model on the same set of terminal tasks, with its 95% confidence interval. The same model in two harnesses shows what the harness adds; the same harness with two models shows what the model adds. Source: tbench.ai.
Every stable release in the last 180 days, one tick each. The leaders ship several times a week; some once-famous names have stopped.
Licence from the project’s own LICENSE file, everything else from its documentation, each fact linked to its source and dated.