Cross-vendor subagents and adversarial checks tame deep-research token burn

A researcher at Quesma set out to map the economics of agentic coding and became a case study in the problem. His first deep-research run exhausted an entire Claude Max 5x plan in 30 minutes: it spun up 111 agents and queued 123 claims for verification, but only 25 were checked before the limit hit, and the synthesis never ran. The cost fix was to stop leaning on one subscription. By extending the claude-mem plugin to share memory across Claude, Codex, and Antigravity, he turned three plans he already pays for into a single pool. A small Bash wrapper lets Claude call the other vendors’ CLIs headless as subagents, so their work draws down Codex and Gemini quotas rather than Claude’s, with automatic fallback when any plan trips a usage limit. Pinning a specific model to each role mattered too — subagents inherit the parent model by default, which was the actual cause of the initial wipeout. The result was roughly ten times more continuous research for no extra spend.