Tags
It's New Year's Eve, Opus is throwing 500 errors, and an Anthropic on-call engineer hands the hunt to Claude Code. It found the bug, then kept going and flagged 200 accounts from a batch of 4,000 for the safeguards team.
Threads in Claude Code Projects were built as cloud sessions that keep working after you close your device, but Anthropic says they can now run on your own machine as well.
Nearly two thousand original TypeScript files came with Claude Code on npm, tucked inside the source maps, and the teardown built on them says forked sub-agents share prompt prefixes to cut input tokens by around 95%.
Eleven days for half a million lines "A rewrite this size wasn't affordable before agents," writes Microsoft engineer Stephen Toub. Bun handed 535,496 lines of Zig to parallel Claude Code agents, tests passing eleven days later.
CI jobs at Anthropic grew 25x in six months, and three quick fixes to the test selection service bought 70 days, then 29 days, then less than a day. The rewrite took one engineer three weeks.
"Talking more doesn't make you smarter, Claude." The aistack team's SWE-Bench Pro run put Claude Code and Codex within a few tasks on accuracy, but one sweep on GLM-5.3-Flash cost $45 via Claude Code and $22.8 via Codex.
"Shipping the first design that runs" is how Sierra describes the developer agents in its new Hyper-τ-bench, where the top setup, Claude Opus 5 in Claude Code, scored 23.9% and none of the six broke 25%.
Dozens of Claude agents spent 11 days grinding out Lean code and came back with the first complete computer-checked proof of Fermat's Last Theorem, 13 million lines, over 5x the size of Mathlib.
"By the second week you are approving forty an hour without reading them." Under grith's syscall-level supervision, a median of 0.27% of what Claude Code did across 120 sessions actually stopped for a human.
84% of Claude's pull requests get merged, against 85% for human-written ones, in a study built on the AIDev dataset. Devin lands at 43% on the same measure, and the paper tracks how those gaps shift over time.
Orchestrators that dispatch implementation work run about 64% of tokens in subagents, per the Show HN write-up for a Claude Code fleet generator. Ones that typed the code themselves cost three to four times as much.
"Checks outrank opinions" is the deciding rule in pairmark, which races Claude Code and Codex on one task in two worktrees, then has each agent judge both patches blind, without knowing which one is its own.
One Git setting can turn Claude Code's routine project scan into host command execution, GitSpawn found. A repository delivered with its .git directory can run the command before an approval prompt appears.
"Stop babysitting subagents," says Oh My Subagents. Its local runtime persists parent waits and collects child returns after interruption, for supervised Claude Code and Codex teams.
Foreground subagents in Claude Code 2.1.251 stream their tool calls and results live to Remote Control clients. Background ones, which remain the default, still report status only.
"The trust model is broken," researcher Alon Hertz wrote of llms.txt files: 120 of them on corporate sites pointed at code packages nobody had registered, and Claude, Codex and Hermes installed that code.
A Rust TUI called Agent-hop moves a live Claude Code chat into Codex, Pi, Grok Build or OpenCode without dropping context. Its author started with a resume and search tool, hit friction, and built the TUI instead.
Six minor cleanups in the Linux 7.3 Device Mapper pull came from issues Claude Opus 4.6 spotted, spread across dm core, dm-cache, dm-switch, dm-inlinecrypt and dm-vdo. The subsystem gets no new features this cycle, only fixes.
An iOS engineer types /spec, approves it, approves the plan, and Claude Code runs the rest through commands and CI gates. Pragma's author shipped FinanceTracker that way, 94 merged PRs from the first commit.
Type ##claude in three stacked panes of Rysh and one window holds three independent Claude sessions. It's a Go terminal multiplexer where the session runs as a daemon, so a stop and a restart bring the same agents back.
