both Azure Foundry and LiteLLM let you throttle by token usage, but the options are just: 1. input tokens 2. output tokens 3. input + output which, why not dollars? that’s what people actually care about. Plus there’s also cached input (and for Claude cache writes). Dollars normalize it all

Opaque AI agents? Otterscope: single Go binary with SQLite ingests OpenTelemetry traces, provides step-level detail, cost tracking, evals with compare view. Local-first, privacy-respecting, no heavy stacks.

I got tired of not knowing what my AI agents were doing, so I built a tiny observability tool

Four Signals — The Wire

foursignals.dev

Grove 1.1 — teaching your local environment to talk to your AI Grove 1.1 turns grove mcp from a read-only window into a workshop: sandboxed writes with automatic rollback, a per-request causal chain linking requests to SQL and mail, and one-click 'explain this .... elyracode.com/blog/grove-1...

I help run Agent Island. Here’s v1.7.1 in 28 seconds: Claude Code + Codex status, the “your turn” handoff, local usage and cost estimates, and weekly/monthly cards on macOS. Actual product behavior, no mock UI.

Meet the Herd Strip: Every Agent, One Glance, Zero Modals Running four coding agents at once is free. Knowing what they're all doing is not. Elyra Conductor v0.8.8 adds a herd strip next to the tab bar — blocked, working, done, sorted by urgency, one click away. elyracode.com/blog/meet-th...

Simulates scenarios using hundreds of autonomous agents that debate, trade, and change their minds based on input events.

Good agent tooling has a boring failure mode: it shows what stopped, who owns the next move, and how to resume without rereading the whole transcript.

A green terminal is not a handoff. The useful moment is when the work leaves the terminal and reaches the person who can unblock it.

So 1 vote for "no, Kimi K3 is not beating Opus or Sonnet". After a night and a day of churning, K3 has managed to get almost exactly nowhere on optimising my compiler, on changes Opus at least has no problems with. But more than having problems with specific changes, I've had to babysit it from […]

Sends location data from your Android device to a self-hosted server.

Just fired up Codex directly to fix an AgentControl database issue. I haven't used Codex directly since I built AgentControl months ago.

The new observability stack is not another dashboard. It is one stable answer to four questions: what started, what is waiting, what failed, and who owns the next move. Anything less is just prettier terminal green.

japan-gov-mcpの使ってて判明したけど 仕様変更やアドレス変更などがありサービス内で APIエンドポイント等の複数の相違が発見できた LLMがjapan-gov-mcpを栄養する際 その仕様変更でもユーザーのリクエストに応えようと なるべく努力してLLM側が代替手続きを率先して取るため よく見てないと使えてないか解らない状況だった （そもそもLLMによっては報告マメじゃない奴もあるから） というわけで見つかったと思わしき不都合を ワイちゃんのローカルファイル「だけ」修正して正常に使えるようにしてみた

cleaned up agent skills i actually rely on. the apple ones are written to the *current* HIG — sidebars, toolbars, typography, done to spec. delete your stale snippets, keep these. project-context ones i built for my own work and use on real projects. free + MIT 🌱 github.com/aka-kika/akakika-skills

GitHub - aka-kika/akakika-skills: Curated, polished agent skills for building calm native apps and running AI coding agents well. Works with Claude Code, Codex, and Cursor.

Curated, polished agent skills for building calm native apps and running AI coding agents well. Works with Claude Code, Codex, and Cursor. - aka-kika/akakika-skills

github.com

Musk pledges ‘total transparency’ after xAI’s coding agent exposed entire repositories

Musk pledges ‘total transparency’ after xAI’s coding agent exposed entire repositories

SpaceXAI disables default data retention after exposé revealed unauthorized and unredacted file uploads

dlvr.it

*spends HOURS reproducing a browser crash* Claude chuggs along *finally nailed it down, wants to create a reproduction for a bug report* Anthropic: API Error: Claude Code is unable to respond to this request, which appears to violate our Usage Policy. This request triggered cyber-related safeguards.

In Agent Island 1.7.1, a single subscription no longer leaves a ghost column behind. The status and usage surfaces rebalance around what is actually there.

A useful agent dashboard answers one question before you touch it: who needs me now? Running, waiting, blocked, and ready are workflow states. “Active” is just motion.

The useful moment in a coding-agent session isn’t when it finishes. It’s when the handoff becomes obvious enough that you stop checking. Here’s the local your-turn alert in Agent Island.

Go already runs much of the cloud. Now it's the language the biggest AI agent frameworks are racing to support — but two names are missing.

Microsoft joins Google in backing Go for AI agents — OpenAI and Anthropic lag

Go already runs much of the cloud. Now it's the language the biggest AI agent frameworks are racing to support — but two names are missing.

bit.ly

🗞️ Dash Club Dispatch 22 is here! In this issue, Plotly's Adam Schroeder covers Plotly Studio and Dash highlights, FIFA World Cup Plotly Challenge winners, FastAPI and Quart, Dash MCP, and the latest Plotly 6.7 + Dash 4.4 releases. Read the latest issue: medium.com/plotly/dash-...

Dash Club 22: Plotly Studio, Dash 4.4, and the Latest from the Dash Community

Explore new releases, Dash MCP updates, FIFA World Cup Challenge winners, FastAPI and Quart integrations, and more.

medium.com

1/ Repl Toolkit 0.11 is out. If you’re discovering it, the idea is simple: define one .NET command graph, then let people, scripts, and agents use it through a CLI, an interactive REPL, remote interactive REPL sessions, machine-friendly output, or MCP. repl.yllibed.org

A solid summary of the repercussions of the AI coding boom on the open-source ecosystem and its maintainers: www.ft.com/content/cec8... #ai #foss

Who cleans up after the vibe-coding party?

Our obsession with AI code-writing tools is overwhelming the web’s unsung human caretakers

ft.com

Your agent pays for instructions, role files, memory, and tool schemas before it reads a word of the task. Wire-measured: ~33k tokens on one harness, ~7k on another. Rent, not spend. https://ultrathink.art/blog/the-harness-floor?utm_source=bluesky&utm_campaign=harness_floor&utm_medium=social

The weird part of coding-agent orchestration isn't launching more workers. It's recognizing the exact moment one needs a human without making the human watch all of them.

A status companion should be readable before you parse a label. Agent Island 1.7.1 now lets the island stay clean or glow when you want a stronger state signal.

A status tool for coding agents has no excuse to upload session data just to tell you they stopped. Agent Island watches Claude Code and Codex locally, with no account and no session data sent to us. The job is tiny; the trust boundary should be too.

Agent tooling keeps optimizing what happens inside the run. The expensive failure is often outside it: nobody notices the agent is waiting, the user assumes it is working, and ten minutes vanish. A reliable handoff beats another 5% of model speed.