On GPT-6 Astra, Reason Machines' harness scored 65.1 against Codex's 62.5 on Artificial Analysis's Coding Agent Index (DeepSWE, Terminal-Bench 4.0, SWE-Atlas-QnA), with an unmodified base harness at 61.4. It cuts tokens via parallel tool runs inside one JavaScript program and a small always-visible tool set, with platform features deferred until code asks for them.
Claude Code's built in subagents stay on Anthropic models; this MIT plugin runs DeepSeek, Gemini (Antigravity), Codex, OpenCode, or Copilot as worker CLIs, each in an isolated git worktree with your verify command, then returns a short verdict and patch for Claude to review. Install with /plugin marketplace add rbinar/cli dispatch ; API keys stay in a local browser form and never pass through Claude…
Piotr Sarnacki found the agent built Rust and Elixir versions made different unprompted calls on backwards compatibility, Redis, and blocking database work, and the Rust one delivered about 1% of notifications under heavy load because of a 256 slot broadcast channel. Rerunning at a constant 100 POSTs/s, he argues dropping lagging clients beats Elixir's unbounded mailboxes, which reached 1.8GB and event delays near 180s…
Ted Haley's free MIT macOS tool indexes code, docs, and past Claude Code/Codex sessions on device and returns the few relevant passages over MCP, cutting lookup tool calls from 3.4 to 2.0 in their Claude Code benchmark. Apple Silicon, macOS 14+; pensieve integrate recommended wires Claude Code, Codex, Cursor, and other MCP clients…
The MIT Codex plugin and local MCP server create or attach Claude Code sessions, hand off scoped implementation in separate worktrees, then review and merge with independent frequency and reasoning knobs. It needs Codex CLI, a working Claude Code setup, Node.js 20+, and Git; Windows and macOS are supported, and Orca is optional…
The environment, not the model, logs every file a terminal agent reads, tracks listed but unopened files, and blocks the first finish while requirements are still open; the authors found about one in five values an agent had already seen never made it into the deliverable. With the same Pi based harness it beat plain direct workspace agents and model kept TODO lists on Workspace Bench, TheAgentCompany, and OfficeQA Pro, lifting Gemini 3.8 Flash's Workspace Bench pass rate from 31.4% to 46.9%…
Spotted something we missed? Start a thread.