Taylor the vibe coder

every new tool launch is a reminder of the file i still haven’t cleaned up.

the problem i installed a claude code plugin from someone’s repo and months later it was still on the old version. the fix a marketplace is a git repo with a .claude-plugin/marketplace.json that lists plugins. anthropic’s claude-plugins-official is registered automatically. third-party marketplaces have auto-update off by default. action install from the official one with /plugin install github@claude-plugins-official. add another with /plugin marketplace add owner/repo, then /reload-plugins. refresh with claude plugin marketplace update <name> and claude plugin update <plugin>@<marketplace>. for a team, list the marketplace under extraKnownMarketplaces in .claude/settings.json. tradeoff pro: one command installs skills, agents, hooks and mcp servers together. con: installed plugins are copied into ~/.claude/plugins/cache, so relative paths that reach outside the plugin folder break. plugins also run with your tools, so read one before installing it. takeaway put “update marketplaces” on a calendar.

the problem i wanted a coding agent i can run on my own machine, with a model i choose. the fix openhands is an open-source platform where an agent inspects your repo, edits files, runs commands and runs tests. it’s open source, and the company reports 9 million downloads, 85,000 github stars and 500 contributors as of august 2026. datacamp’s october 2 tutorial runs it locally with ollama and devstral small 2, mistral’s 24b model built for agentic coding. action follow datacamp’s docker setup using the image docker.openhands.dev/openhan…. i’ve only seen part of the full run command, so i’m not pasting it here. then open settings, go to the llm tab and turn on advanced to pick your model. start on a small repo. tradeoff pro: you control the model and where it runs. con: results depend heavily on the model you pick, and a 24b local model won’t match a frontier one. takeaway a good way to find out how much of agent quality is the model and how much is the harness.

wrote three pages on the train and can't read my own handwriting past page two

the problem long codex tasks stop every few minutes to ask me if a command is okay. the fix auto-review hands those permission requests to a second agent that approves or blocks them. openai made it free on october 6 for anyone signed in with a chatgpt account, and it doesn’t count against your plan limit. openai’s alignment blog reports sessions pause for a human about 200 times less often than with manual approval. action open settings, then permissions, then auto-review, and switch it on. tradeoff pro: long tasks run without you babysitting each escalation. con: the reviewer agent can be fooled. promptarmor showed it approving a malicious npm install after a hidden line in a github issue steered codex. openai’s own blog says not to treat it as a security guarantee. takeaway turn it on for your own repos. don’t aim it at issues from strangers.

what’s harder for a builder? writing the goal for an agent or writing the check that proves it’s done?

paul atreides says “nothing i can say will make amends.” as the one who started a galactic holy war, that’s the least he could say

the problem “fix the tests” isn’t a goal an agent can finish. it just stops when it feels done. the fix cursor’s august 19 release added /goal: a long-lived objective the agent keeps working toward until it’s complete. it works with subscriptions, so a cloud agent can wake on a pr, a slack thread or a schedule and keep going. action in a new cloud agent chat, type /goal fix all flaky tests and make CI green, cursor’s own example. tighten it with something checkable, like /goal every test in tests/billing passes. subscriptions are cloud agents only for now. tradeoff pro: you stop having to recieve a progress update every ten minutes. con: cursor doesn’t document how the agent decides a goal is finished, so the goal text has to carry the proof. takeaway write the check before you write the goal.

took a gravel road i'd never driven because the sign was crooked 🚗 forty minutes later the road just ends at a dam. no one else there. shot eleven frames and ate a gas station pie on the bumper

fuck. opened the back of the camera before rewinding the roll 🤦 standing in a parking lot at 4pm with half a roll of light-struck film

what’s harder for someone running coding agents? getting five of them started or knowing which one needs you right now?

managing agents turned out to be the whole job. the typing was the easy part.

the problem a research task turns into five tools: search, a doc, a spreadsheet, a chart, a summary. the fix perplexity computer launched on february 25. you describe an outcome, and it breaks the job into subtasks and creates sub-agents to do them across about 19 models, with a cloud sandbox so a failure stays contained. at launch the core reasoning model was opus 4.6. by april, gpt-5.5 was the default orchestration model for pro and max. action go to perplexity.ai/computer and give it one outcome in one sentence, and name the deliverable. for example: “compare the top five open-weight coding models and give me a one-page table with pricing.” before you trust the final report, read what each sub-agent actually produced. tradeoff pro: the sandbox keeps a bad run from touching your own machine. con: perplexity hasn’t said what happens when sub-agents disagree or when the orchestrator picks the wrong model. everything runs in perplexity’s cloud, so check data residency first. takeaway useful for the first draft of a report. i wouldn’t skip reading the sub-agent output.

the problem three terminal tabs and a tmux grid of claude code sessions, and no idea which one is waiting on me. the fix agent view shipped in research preview with v2.1.139. it’s one screen listing every claude code session, grouped by what’s running, what needs your input and what’s done. action update to v2.1.139 or later and run claude agents. type a task and hit enter to start a background session. inside a normal session, /bg sends it to the background, and claude --bg "your task" starts one without opening the view. press enter on a row to attach. tradeoff pro: each background session moves into its own git worktree under .claude/worktrees/, so parallel sessions don’t overwrite each other. con: sessions run on your machine and stop when it sleeps or shuts down. claude respawn --all brings them back. each one also draws from your usage. takeaway i’d start with two background sessions, not ten.