vibehacker
News
mtlynch.io (via Hacker News) ·

"Why are coding agents so dumb?": mtlynch argues the harness, not the model, is now the bottleneck

A daily OpenCode and Claude Code user says agents still run parallel subtasks one at a time, never hand gruntwork to cheaper models, write hard-to-read plans, stall overnight on trivial questions, and lack real OS-level sandboxes. He lists table stakes he wants out of the box, including per-task model routing, deterministic filesystem and network boundaries, and an AFK mode that decides after 30 minutes without a reply.

Read at mtlynch.io

More news

View all

Karpathy: the gap in how people understand LLM capability is now a sharp funnel, and it's widening

By his napkin math, about 6B people have barely touched LLMs, 1 to 2B use free tier chatbots, roughly 20M see agents finish weeks long math and code projects from a prompt, and about 5,000 inside labs see swarms of agents working for weeks. He blames scale of ambition, capability peaking in verifiable domains like code, and access tiers, and warns that human review is starting to fall behind…

Andrej Karpathy on X (via Hacker News)

Harvard study of 718 firms: coding agents add 30% more code but no measurable rise in resolved issues

Using Jellyfish data on about 300 million work events, Fiona Chen and James Stratton find agent adoption raised lines, commits, and PRs by 20 to 30% while completed Jira issues and epics showed no statistically significant gain. They point to review as the bottleneck: PR submission to merge time rose 49% (3.45 days), change requests nearly doubled, and comments per PR rose 35%…

Harvard working paper (Chen & Stratton)

bigarrow lets Claude Code and Codex draw arrows on your Mac screen when they need you to click something

The MIT Swift CLI and agent skill puts a labeled, click through arrow, box, or ring over any window (an Allow button, a 2FA field, the right Chrome tab) without taking focus, and it never clicks, types, or captures the screen itself. Each arrow removes itself on a timer or when the agent exits, installs with brew install franzenzenhofer/tap/bigarrow , and the Show HN passed 400 points…

GitHub / franzenzenhofer (via Show HN)

Agent Swarm 1.169 adds sub-workflows and lets you click page elements to send feedback to the lead agent

The MIT, self hosted system where a lead agent hands work to Claude Code, Codex, Cursor, Amp and other workers in Docker now has a workflow node that runs a child workflow and waits for it, plus an overlay that sends comments on elements of agent built pages straight to the lead. Version 1.169.0 (Oct 11) also adds an optional gate script that skips quiet heartbeat checks and nightly eval runs with regression checks…

GitHub / desplega-ai/agent-swarm

Agent Standup turns a Claude Code session into a live stand-up call, at zero token cost

The free, open source tool reads the session files Claude Code already writes, so the lead agent gives orders while sub agents join, share their screen with real commands and diffs, and report back, with a who asked whom tree, context gauges, and replays. It runs as a VS Code extension (VSIX for Cursor and Windsurf), via npx @double geste/agent standup open , or as an unsigned desktop app, and its author calls it "mostly for fun.…

Agent Standup (via Hacker News)

Spotted something we missed? Start a thread.