cua-speedrun benchmarks computer-use agents on speed and cost
A pip-install eval runs agents in matched sandboxes, ranked by time and cost. Also: turbopuffer demotes the vector index; Etymon unifies agent configs.

Copy markdown
A fair, fast way to score computer-use agents
cua-speedrun runs your agent and its desktop in separate Modal sandboxes on identical hardware, timing each task from instruction to completion and logging cost — not just pass/fail. Its 50-task OSWorld subset correlates 0.98 with the full 295-task run while cutting eval time 84.9%, and a Fast I/O mode drops per-screenshot delay from 2-3s to as low as 2ms. Run `pip install cua-speedrun`, point it at an agent.py, and compare your build on OSWorld, OSWorld 2.0, CUA-World, and MyPCBench.
turbopuffer demotes the vector index
In a post bluntly titled “RIP, vector database,” turbopuffer says its coming v3 rebuild makes the ANN index just another secondary index rather than the primary one — ending storage duplication for multi-vector docs and write amplification on rebalance. The service already runs 1T+ documents at 10M+ writes/s and 25k+ QPS, and its earlier full-text v2 rebuild shipped a 10x smaller index with queries up to 20x faster. If you run hybrid text-plus-vector search, your backend is about to get faster — though v3 is still pre-GA.
Etymon: one config for every coding agent
This MIT-licensed CLI collapses CLAUDE.md, AGENTS.md, .cursor/rules, .codex/config.toml, and .mcp.json into a single etymon.toml plus lockfile. Run `etymon sync --harness claude,codex,opencode` and it regenerates each tool's native files from that one source, so a repo stays portable across harnesses and you stop hand-syncing the same rules into five places.