vDesktopAI Masterclass
See the plans
← How we build

The Agent Software Signal

The tools and benchmarks we build with, tracked in the open.

vDesktop builds software with AI agents, so we watch the agent-software landscape closely — the benchmarks, the models, the frameworks, the research. This page is that watch, made public: the same figures our build decisions rest on, each one showing its source and when it was last checked. We do not run these benchmarks and we do not own these numbers — we poll public sources on a timer and republish what they say, unchanged.

We report these numbers. We don’t run them.

Every figure links to the source it came from and shows when we last checked it. If a timestamp is old, the number is old — read it that way. The reason this page exists is the same reason we publish our own build record: you should be able to see where a claim came from before you rely on it.

How well agents actually resolve real issues.

The published SWE-bench and SWE-bench Verified leaderboards — how well coding agents close real, previously-solved GitHub issues. Reported from the public leaderboard.

SWE-bench Verified ·
Top 10 by resolve rate
  • Loading the current leaderboard…

A high score means an agent can close curated issues, not that it will close yours. We read it to decide what is worth trialling, never what to trust unseen — which is why everything we ship still goes through the checks on our Practice page.

The models we build with, and what they cost.

Current frontier and leading open models with their context windows and published prices. Sources: each vendor’s own model and pricing page, hand-checked.

Price and context window change what is worth building more than raw capability does — a job that was uneconomic last quarter can become routine on a cheaper, longer-context model. We rebuild what runs where around these numbers, so we track them rather than assume them.

Which foundations are still being maintained.

Development activity on the main open agent frameworks — stars, latest release, and commit velocity — straight from the GitHub API.

Momentum is a maintenance signal, not a quality one, but a stalled repository is a liability you inherit the day you build on it. We watch releases and commit cadence to know which foundations are still being fixed underneath us.

What the field is still working out.

The most recent arXiv papers on autonomous and coding agents, filtered from cs.SE and cs.AI.

    Most of this never reaches production, and reading it early is how you tell which problems the field still considers open. A technique here is a lead to test, not a feature to promise a client.

    This is the same care we build with.

    If reading the sources before trusting a number is how you want your software made, tell us what you need built.

    Tell us what you want built Read the build record

    This page reports public sources and owns none of the numbers on it. Each figure shows when it was last checked; if those times are old, treat the page as old.