Stateless Language Agents cut token usage by over 84%

Stateless Language Agents (SLA) move durable research history into an external harness so each agent receives only concise evidence for its next step. The approach matches the strongest kernel baseline's final performance while using over 84% fewer tokens, with the Advisor coordinator consuming under 0.6% of the budget.

• SLA achieves the best final result on every task with over 84% fewer tokens
• Savings hold at budgets up to one billion tokens across three domains
• The Advisor coordinator consumes less than 0.6% of tokens
• Evaluation is limited to three engineered domains