
The Workbench: Build Your Own AI Switchboard
A personal AI switchboard you can build this weekend: one router that sends routine work to a cheap model and hard reasoning to a frontier one, with a spend cap you set.
Tag
23 entries tagged AI Tools · 19 wires · 4 terms.

A personal AI switchboard you can build this weekend: one router that sends routine work to a cheap model and hard reasoning to a frontier one, with a spend cap you set.

Three tools that stop making you go to a chat window and bring the agent to where you already work: Framer's canvas, ACME.BOT's interview, and Deck's inbox.

Three releases that move the operator off rented AI and onto hardware they own: Colibri runs a frontier model on a plain laptop, BrowserOS keeps your browsing local, and Inkling hands you the weights.

OpenAI retires its year-old Atlas browser, merges ChatGPT, Work, and Codex into one desktop app, and debuts GPT-Live voice: the churn even the leaders live inside.

Meta's Muse Image launched July 7 on Instagram and WhatsApp, free, with AI editing access to any public Instagram photo. Claude Fable 5 moved to $50 per million output tokens today. Two Anthropic policies took effect at once as Meta put AI image editing in the pocket of three billion users.

Three tools for the week you stopped being locked to one provider's meter: OpenCode's model-agnostic agent, Grok Build's /goal mode, and Kombai's Figma-to-code specialist.

Google shipped Gemini 2.5 Pro with Deep Think and a 2M-token window on June 22. Claude Fable 5 remains offline under a government export ban. The accessible frontier shifted.

Z.ai opens GLM-5.2 for long-horizon coding, ChatGPT graduates scheduled tasks to general availability, Google forces every Gemini CLI user onto Antigravity. Three tools that move the unit of work from the prompt to the standing instruction.

Claude Fable 5 lets the model scale its own thinking, Lathe flips the LLM script from do-it-for-me to teach-me, Apple opens the on-device model and refuses to pick a winner. Three tools that change who the work is shaping.

Zoom ships an agentic teammate, Microsoft puts Scout in the OS, NVIDIA opens a 1M-context model built for long-running agents. Three tools that move AI from prompt to coworker.

Anthropic ships dynamic workflows and a real-time Claude Code security plugin, Bluedot routes Apple Watch recordings into Claude. Three tools that put rails around the model.

Cursor ships its own coding model, Google relaunches Antigravity as an agent-first platform, and Anthropic adds sandboxes and MCP tunnels for Claude. Three editors collapse into the agent.

OpenAI ships Daybreak for the security pipeline, Anthropic launches Claude for Small Business with QuickBooks built in, and open-source OpenHuman builds a local Memory Tree. Three tools, one pattern.

SpaceX locks a $60 billion option on Cursor with 200,000-GPU Colossus compute access. GPT-5.5 surfaces briefly in OpenAI Codex. Anthropic's Claude Design erases $700 million from Figma's market cap in 48 hours.

Three tools that hand the craft over to an agent this week, and the warning that should land alongside them.

Three tools worth your weekend: Karpathy's Autoresearch, Microsoft's agent framework, and the open-source Claude Code alternative nobody is talking about.

OpenAI shut down Sora and walked away from a $1B Disney deal in the same week. What the pruning actually signals, plus a practical look at Claude Code for non-technical builders.

ChatGPT rolls out ads to free users, Anthropic ships Claude Code Channels, Musk unveils Terafab, and what all of it means for builders who aren't paying attention to the right things.

Anthropic surveyed 81k users using Claude itself as interviewer, Cursor shipped its own coding model at a fraction of frontier prices, and Google turned AI Studio into a full-stack builder. Here's what matters for operators.
The pattern in which an agentic AI system runs its own pass of checks against its own output before it asks the operator to look.
A JavaScript orchestration script that Claude writes and a runtime executes to fan a complex task across tens to hundreds of parallel subagents, keeping intermediate state outside the conversation.
The practice of deliberately routing different phases of an AI task to different models based on the price-to-quality ratio each model offers for that specific phase.
A saved prompt paired with a cron-style or event trigger that lets a chatbot or terminal agent wake itself up, run a workflow, and surface the result without a human pressing send.