@interjc: Over the past few days, I used my own subagent-team skill to lead an "AI team" in development — burning through the weekly quotas of several Claude/Codex/Gemini accounts plus hundreds of dollars in API costs on a major overhaul of a monorepo project. I pulled out all the delegation records and did a retrospective:

Over the past few days, I used my own subagent-team skill to lead an "AI team" in development I burned through the weekly quotas of several Claude/Codex/Gemini accounts plus hundreds of dollars in API costs on a major overhaul of a monorepo project I pulled out all the delegation records and did a retrospective: The main agent was Claude Code, the driver was Opus 5.5 High, and the downstream tools were codex / agy and opencode (connected to the Claude API and a codex account) Over 300 internal subagents and over 200 external agent tasks (Codex, agy, OpenCode) The project is no small thing: two mobile apps and around eight Cloudflare Workers services. We did SDK upgrades, consolidated authentication, prepaid credits billing, an AI gateway, and an interface redesign. The main session had Opus 5.5 acting as director, responsible for task breakdown, writing contracts, acceptance, and commits — everything else was delegated out The most surprising finding: the most cost-effective use of external agents isn't writing code, it's reviewing. I had Codex's gpt-6.1-sol with high effort do read-only reviews of billing and auth code, and all 7 times it found real issues: charging a fee without returning results, double-charging, and reading other people's data via an internal token. Having a model from a different vendor look at the code fills in exactly the blind spots of your own model. Internal reviews were also worthwhile: out of 72 reviews, 21 raised blocking issues, nearly all confirmed by the implementer, with only 1 false positive The second finding: the amount of rework depends on task risk, not model tier. Code written by Opus and Sonnet both required rework after review at a 67% rate. Whenever money, auth, or table schemas are involved, no matter who wrote it, it has to go through several rounds of review — so rather than switching to a more expensive model, it's better to build review rounds into the schedule from the start Higher effort isn't always better either: Opus at xhigh took a median of 19 minutes per task at $3.70–11.50, and reviews still caught the issues that needed fixing every time; Sonnet at medium for small, clear changes took about a minute, costing anywhere from a few cents to a little over a dollar There's also an easily overlooked role: the "coordinator." Every external task was paired with a Haiku low subagent responsible for dispatching work, waiting for results, running acceptance checks, and verifying the scope of changes — the vast majority completed successfully External agents often can't run tests due to sandbox restrictions, so this layer is what actually does the real gating. Every external task failure was environmental (effort settings getting reset, sandbox denials, timeouts) — none of them modified files outside the intended scope, and not a single result needed to be thrown out and redone To summarize my division of labor: Low-tier models as coordinators; Mid-tier models for well-defined small changes and research; Strong models for risky implementation; Top-tier models and strong models from another vendor for reviews;