At Quesma, researching AI agent tokenomics, the author's initial /deep-research run burned the entire Claude Max 5x limit in 30 minutes, launching 111 agents but verifying only 25 claims before hitting the cap with no final synthesis.

The fix leveraged three existing subscriptions—Claude, Codex, and Antigravity—with a locally extended claude-mem plugin for shared memory. A model-orchestration pattern assigned roles: Claude Fable 5 plans, Sonnet 5 finds, Opus 4.8 verifies, Haiku 4.5 handles extraction, Codex (GPT-5.5) runs tools, and Antigravity (Gemini 3.1 Pro) provides a second opinion. A Bash wrapper calls other CLIs as headless subagents, falling back to Claude models when limits hit.

This yielded roughly 10x longer research sessions (hours vs. minutes) at no extra cost. Hallucinations were reduced by requiring independent verification, primary-source URLs, and quoted evidence for every claim. The /deep-research tool now runs last, deepening pre-verified findings; a recent run used 61 agents in 22 minutes versus the original 111 agents in 30 minutes.

Validated notes flow into an Obsidian LLM wiki inspired by Karpathy's pattern, now holding hundreds of atomic notes on pricing, tools, and benchmarks. Automated checks alone aren't sufficient—one rule once rejected Headroom, a 56k-star project.