Quesma researchers developed a cost-effective, multi-model agentic pipeline that optimizes token usage and improves data reliability by orchestrating various AI subscriptions through a shared memory framework.

Key Points

• The system uses a tiered model approach: Claude Sonnet 5 for searching, Claude Opus 4.8 for verification, and Claude Fable 5 for high-level planning.
• A custom Bash script integrates Codex and Antigravity as headless subagents, allowing the pipeline to leverage multiple existing subscriptions simultaneously.
• Implementing strict verification rules—such as requiring primary source quotes and URLs—significantly reduced AI hallucinations during the research process.
• Moving the resource-intensive /deep-research tool to the final step of the pipeline reduced token consumption while increasing the quality of synthesized findings.
• Research revealed that poor harness configuration and inefficient context compaction can increase token costs by up to 100% or more.

Why it Matters

This hybrid approach demonstrates that businesses can significantly lower AI operational costs and improve output accuracy by moving away from single-model reliance. By implementing human-in-the-loop verification and strategic model orchestration, organizations can build reliable knowledge bases without exceeding budget limits.