A full-time agent user's daily stack, with Jev for quick judgments
The author shares the agent setup they use every day: Opus 5.5 as the main driver and Jev for fast yes/no judgments at specific points in the workflow.
I've spent nearly every day since February full-time with AI agents and harnesses, writing code and managing my life. I've tried so many things. Here's my latest set-up which is working super well for me. Opus 5.5 is my daily driver. I move between the Claude Code and Codex desktop apps depending on which has the best models and features at the time, and from whichever one I'm in, I call in the others. Right now I'm in Claude Code, and it's the best experience I've had in a while. Would love to hear how you all do it, and ways to improve my system. Models I call in: — Codex (GPT-6 Astra, GPT-6.1 Sol): independent code reviews, second opinions, longer builds that run on their own — DeepSeek V4.1 Flash: small, self-contained coding jobs in a sandbox — Jev by TypeSafe: fast yes/no judgments at specific points in the workflow (below) — Gemini: critiquing video frames, Nano Banana Pro image edits, Lyria music Research on X: — Grok Bot: I message a Grok bot on X straight from Claude Code, and its research comes back to the session a few minutes later — xAI API (Grok 4.7 with X and web search): when I need an answer right away — X Search, an X API plugin I made for Claude Code and Codex: exact posts, threads and metrics, to check what Grok cites. http://github.com/ClawRyderz/x-search-mcp-plugin — My X bookmarks: a weekly script reads a bookmark folder so my agents see what I've been saving Running locally: — Fluid Voice on my laptop: dictation with Parakeet speech recognition and a small cleanup model, all on-device — A Mac Studio at home: Qwen 3.8 27B for text, DeepSeek-OCR for documents, and all our ffmpeg renders Film and media: — OpenAI GPT Image 2.5: my main image generator — Midjourney V8.2: the look of a film and character references — Nano Banana Pro: editing stills and first frames — ElevenLabs: voices, voice design, sound effects, music, and timing alignment — Kling 3, MiniMax, LTX-2.5, Seedance 2.5, Wan, Runway Act-Two, OmniHuman (mostly through http://fal.ai): video shots — http://sync.so, VEED and LatentSync: lip-sync, scored automatically so bad takes drop out before I see them — Tools we wrote: face consistency checks, tracked screen inserts, a lip-sync nudger, contact sheets — Topaz Starlight: upscaling, last, on locked shots only — DaVinci Resolve Studio: editing and color, which Claude drives directly — HyperFrames: motion graphics written as code, with a critic agent — Blender: 3D renders on a home server How work flows (loops and graphs): — Every task starts with a "done when" list a program can check — Claude splits the work and hands bounded pieces to Codex, DeepSeek or subagents, each in its own git worktree — When a worker reports back, Jev compares its report with the done-when list and points to what to check first. The real checks still run. — A loop that makes no progress after two tries stops, and the plan gets rethought instead of retried — How much checking a change gets depends on how hard it is to undo: routine changes merge once CI passes, critical ones wait for an independent review, live systems follow a written protocol — Every real incident becomes a small automatic check, and a weekly retro reads my past week's sessions and bookmarks and suggests changes Olympus, which I'm building: A private second brain that lives on my computer and works with any AI. It's how my agents search my email, files, messages and bookmarks. Everything is stored and searched on my own machine, each item is marked Personal, Private or Secret, and I decide what each model sees. Small built-in models answer on my computer first: Qwen3.5 for private answers, EmbeddingGemma 2 for search, Qwen3-ASR for voice notes. When I want a bigger model, it can use Venice's private models or ask anonymously over Tor. Open source, no account, and it works with Claude, ChatGPT, Codex, Grok and OpenClaw. 1.0 just shipped: http://olympusplugin.ai Everything else: — Castor: an always-on assistant (OpenClaw) on a home Linux server that I message from my phone — That server also runs builds, renders and agents that work while I'm away — Remote Control: I check on and steer any session from my phone — Agents running in parallel can see which files the others are editing — One instructions file and one skills library shared by Claude Code and Codex — GitHub: every change goes through a pull request and CI