Coding agents

How I'd choose between Claude Max, ChatGPT Pro, team seats and API billing after DevDay, with the limits and cost comparisons that matter.

On Artificial Analysis's Terminal-Bench runs, Opus 5.5 scores the same at xhigh and max but takes five times as long to start answering. My levels, and why.

I run CLIProxyAPI on a Raspberry Pi at home so Claude Code and Codex on every machine can use my accounts without logging in and out. Here are the account rules, the cache fix it needs and the setup over Tailscale.

Run T3 Code on one MacBook with two Claude accounts and four models, reach it from your phone over Tailscale, and cut each agent's background processes so more of them fit in memory.

I've put the main ways of giving coding agents a browser through months of daily use. The ones that hold up give each agent its own tabs or its own browser.

I switched from Handy to VoiceInk for dictating prompts to coding agents. It feels faster, gives me more control and handles corrections better. Handy is still a good option.

I use Humanizer when a message goes out as me and Unslop for everything else. Here's a real example of each, and the rule I gave my agents.

An energy integration in my Home Assistant had been broken for months before I noticed. Now my servers, backups and house report success, so I hear when something goes quiet.

I have AI judges blind-score my agents' work against a rubric, round after round. Here's the loop, a run that climbed from 7.5 to 8.0, and where judges mislead.