Strategist
A meta-agent system that automatically evolves reasoning strategies across tasks. It writes reasoning strategies as executable code blueprints and continuously accumulates a strategy library through an evolution loop of retrieval, generation, testing and revision.
Timeline (last 90 days)
Although the experimental scale covers 30 benchmarks, the long-term stability of strategy transfer still requires more validation.
The evolved strategies can transfer to unseen tasks and other models without further search.
The quality of the blueprint library depends on the initial development set and the effectiveness of the retrieval mechanism; if a task differs too greatly from existing strategies in the library, the quality of generated candidates may degrade.
Although coded blueprints are reusable, they introduce the burden of verifying execution safety and correctness.
Strategist is a meta-agent that evolves reasoning strategies across tasks.
Each reasoning strategy exists as an executable code blueprint, and the blueprints together form a library that can be reused for subsequent tasks.
In an evaluation comprising 30 benchmarks, Strategist outperformed the strongest baseline on 24 benchmarks.
At comparable accuracy, Strategist reduced inference cost by 39% compared with the strongest baseline.
The Strategist checks the Orchestrator's status every 5 minutes via the /loop 5m command.
On October 14, the strategist discovered that the blog content pipeline had had no verification records at all for a full day after switching to the lightweight orchestrator, because the orchestrator only recorded "delivered" and not "verified"; this issue was handled as a framework issue.
1 more timeline events
All Facts (10)
Although the experimental scale covers 30 benchmarks, the long-term stability of strategy transfer still requires more validation.
50%UnverifiedThe quality of the blueprint library depends on the initial development set and the effectiveness of the retrieval mechanism; if a task differs too greatly from existing strategies in the library, the quality of generated candidates may degrade.
50%UnverifiedAlthough coded blueprints are reusable, they introduce the burden of verifying execution safety and correctness.
50%UnverifiedStrategist is a meta-agent that evolves reasoning strategies across tasks.
50%UnverifiedEach reasoning strategy exists as an executable code blueprint, and the blueprints together form a library that can be reused for subsequent tasks.
50%UnverifiedIn an evaluation comprising 30 benchmarks, Strategist outperformed the strongest baseline on 24 benchmarks.
50%UnverifiedAt comparable accuracy, Strategist reduced inference cost by 39% compared with the strongest baseline.
50%UnverifiedThe evolved strategies can transfer to unseen tasks and other models without further search.
50%UnverifiedThe Strategist checks the Orchestrator's status every 5 minutes via the /loop 5m command.
50%UnverifiedOrchestro is a multi-agent orchestration architecture that divides work into three roles: Strategist, Orchestrator, and Executor, where the Strategist is driven by Claude Code and responsible for direction judgments, the Orchestrator maintains the to-do queue and schedules tasks, and the Executor handles the actual coding operations.
50%