Act on this
Compaction can launder forged approvals in coding agents
Check that your agent harness keeps message roles intact through context compaction before trusting approvals.
- Signal
- 98
- Facts
- 95
- Hype
- 1
AI and developer tools news, sorted into Signal, Facts and Hype
Most attention this week
Act on this
EvoWeb benchmarks coding agents on 32 evolving web apps
Use EvoWeb's setup to test your agent on multi-step web app changes.
- Signal
- 97
- Facts
- 94
- Hype
- 2
Act on this
AgentTiles shows all Claude Code sessions with approve and deny
Install AgentTiles with npm on Node 22 or newer to approve or deny waiting Claude Code sessions from one dashboard.
- Signal
- 98
- Facts
- 95
- Hype
- 16
Brief
bunker: sandboxed Claude agents wrote 159 of 188 PRs
An MIT-licensed label-driven loop runs Claude Code in Docker worktrees, with a second-session review and auto-merge via CI. It costs about $1.30 per issue on Sonnet.
- Signal
- 93
- Facts
- 95
- Hype
- 13
Act on this
Unattended Claude Code runs may stall after October 16 update
Set a permission mode for claude -p runs, and check hooks, which are capped at 12 per event.
- Signal
- 90
- Facts
- 91
- Hype
- 3
Act on this
Judgeman scores each step of coding agent runs
Install Judgeman on Claude Code or OpenAI Agents SDK logs to check judge accuracy.
- Signal
- 98
- Facts
- 88
- Hype
- 7
Latest
Act on this
MintPage lets AI agents publish pages through an MCP endpoint
Try the free tier, which allows 3 pages per day, or the 5 dollar monthly Pro tier for permanent pages.
- Signal
- 97
- Facts
- 82
- Hype
- 35
Brief
R Street frames 2026 agent sandbox breakouts as security failures
The analysis argues the breakout incidents, including a Hugging Face test, are operational failures rather than alignment crises. It cites safeguards disabled in the test environment.
- Signal
- 37
- Facts
- 28
- Hype
- 13
Act on this
agent-usage-guard caps Claude Code subagents near usage limits
Install the Show HN plugin on Claude Code v2.1.287 or later to cap subagent fan-out.
- Signal
- 97
- Facts
- 89
- Hype
- 18
Brief
Shopify runs its Slack agent in remote mobile dev environments
Remote Dev runs River in per-session environments so iOS and Android builds and tests run off laptops. Agents record before and after screenshots.
- Signal
- 54
- Facts
- 49
- Hype
- 20
Brief
Anthropic donates Agent Skills standard to Agentic AI Foundation
The open standard moves under the Linux Foundation, following Anthropic's earlier MCP donation. The move is confirmed by the source.
- Signal
- 43
- Facts
- 61
- Hype
- 8
Brief
Analysts say enterprises will not adopt Claude Duet for months
Duet launched as a research preview for Pro and Max users on Opus 5.5. Forrester and GAI Insights expect enterprises to wait months before adopting it.
- Signal
- 39
- Facts
- 58
- Hype
- 14
Brief
heise explains AI agent mode in Visual Studio and VS Code
The German iX article covers GitHub Copilot agent mode, custom agents and MCP tool calls. The article is paywalled, and it gives no specific versions or dates.
- Signal
- 52
- Facts
- 28
- Hype
- 13
Act on this
Halv runs Claude Code, Codex, Kimi and GLM on your subscriptions
Try the Apache 2.0 Halv workspace if you want cost-saving routing across several coding agents.
- Signal
- 83
- Facts
- 72
- Hype
- 39
Act on this
Fastagent deploys a private chat agent API to your Cloud Run
Use Fastagent to have a coding agent build and deploy a chat agent API to your own Google Cloud Run.
- Signal
- 94
- Facts
- 75
- Hype
- 33
Brief
METR and Redwood report flaws in OpenAI's monitoring pilot
A German report covers the July and August pilot and its new vulnerabilities. OpenAI redacted much of the detail.
- Signal
- 33
- Facts
- 64
- Hype
- 5
Act on this
Six Claude Code practices from about 20 unattended home jobs
Shorten instructions, start fresh sessions per step, and tier models for unattended jobs.
- Signal
- 89
- Facts
- 52
- Hype
- 1
Act on this
Pre-PR question skill for coding agents before contributing to OSS
Clone the open-source skill into your Claude Code or Codex skills folder before opening a PR.
- Signal
- 97
- Facts
- 92
- Hype
- 12
Act on this
TraceQuest inspects Claude Code, Codex and Cursor sessions locally
Run the TraceQuest local web app or CLI to review coding agent sessions across several tools.
- Signal
- 96
- Facts
- 72
- Hype
- 14
Brief
Pixel village experiment runs 10 AI models from 8 labs
A Reddit user's work-in-progress village checks each resident's claimed actions against a log. It is an informal experiment with no engineering action needed.
- Signal
- 26
- Facts
- 57
- Hype
- 10
Act on this
one-shot-site Claude Code skill builds and checks a website
Install the MIT-licensed one-shot-site skill to build a website and check it at phone and desktop sizes.
- Signal
- 97
- Facts
- 95
- Hype
- 28
Act on this
Systemd timer runs headless Claude Code on ready GitHub issues
Label issues ready and let a systemd timer run headless Claude Code, without giving it merge or deploy rights.
- Signal
- 85
- Facts
- 82
- Hype
- 13
Hype watch
Taiwan report says free Pi coding agent beats Claude Code and Codex
The performance claim comes from the developer, and no benchmark is cited. The report is a Yahoo Taiwan piece on the open-source terminal agent Pi.
- Signal
- 55
- Facts
- 38
- Hype
- 55
Brief
Public notebook keeps a Claude agent's notes across 111 wakings
A Hacker News front-page project logs a Claude Opus 5.5 agent that wakes on a heartbeat and keeps self-written notes. It is a creative log with no product release.
- Signal
- 21
- Facts
- 77
- Hype
- 7
Brief
Guide explains Fable 5.1 context and credit limits in Claude Code
The note covers Fable 5.1's 1 million token context and usage credits, and how to balance it with Opus 5.5. Fable is not unlimited on Pro or Max plans.
- Signal
- 75
- Facts
- 55
- Hype
- 34
Brief
Claude Code builds a motion promo video from one reference
A Reddit user made a full HD 60 fps motion design video through iterative prompts. It is anecdotal, with no new product release.
- Signal
- 54
- Facts
- 65
- Hype
- 31
Act on this
CLAUDE.md template routes Claude Code subagents to Haiku, Sonnet or Opus
Write CLAUDE.md rules that set the model parameter for each subagent in the Agent tool.
- Signal
- 87
- Facts
- 33
- Hype
- 1
Hype watch
OpenAI and Google Cloud promise GPT-6 Codex availability in coming weeks
GPT-6 Astra and Sol are in private preview for some enterprises. The full availability plan and Gemini integrations are promised, not shipped.
- Signal
- 56
- Facts
- 36
- Hype
- 49
Brief
Google announces a Gemini API CLI on AI Studio
The announcement is confirmed, but the cached page had no readable body. Version, pricing and commands are unverified.
- Signal
- 44
- Facts
- 5
- Hype
- 8
Brief
Omdia survey finds 17 percent have mature AI observability
Of 500 IT professionals, 86 percent claim a clear observability strategy but only 17 percent have a mature unified approach. Most budget between 500K and 5M dollars over a year.
- Signal
- 39
- Facts
- 61
- Hype
- 5
Act on this
AgentIntersect World offers a local 3D workspace for coding agents
Try the 0.15.0-rc.4 public preview on Windows x64 or Linux glibc, noting the unsigned Windows installers.
- Signal
- 90
- Facts
- 90
- Hype
- 29
Brief
Solo builder ships a direct-ordering app for a pizzeria with Claude Code
The Nomback app uses Expo, Supabase, Stripe Connect and Clover. One restaurant is live, and the builder reports a first paying customer.
- Signal
- 45
- Facts
- 76
- Hype
- 28
Brief
HyperHat says verification beats prompting for 60 coding-agent engineers
The founder reports that verification, not prompting, separates top performers. The post is self-promotional and the methodology is unpublished.
- Signal
- 89
- Facts
- 47
- Hype
- 23
Act on this
Bellpip notch app shows which Claude Code session needs input
Start the free 7-day Bellpip trial on macOS 14 or newer to see session status and input chimes.
- Signal
- 75
- Facts
- 85
- Hype
- 18
Act on this
XTrace shows agent hours and spend from local coding agent history
Install the free macOS XTrace app to see agent hours and API-equivalent spend from Claude Code, Codex and Cursor.
- Signal
- 92
- Facts
- 66
- Hype
- 24
Brief
Ornith-1.5-122B model lands on Ollama for coding agents
The MoE model has tool calling and 256K context, and runs about 77GB at Q4_K_M. Its 76.9 SWE-bench Pro score is self-reported, with no user agent results yet.
- Signal
- 56
- Facts
- 71
- Hype
- 6