While coding agents have reached a relatively mature stage, domain-specific agents are lagging far behind; for instance, using Codex to play games yields poor results.
The Spire agent addresses the challenge of maintaining consistency across long-horizon tasks by delegating deterministic actions to domain-specific tools, while the LLM handles logical tasks requiring reasoning and acts as the "glue" connecting the components.
Currently, about 40% of runs reach Act 3, and approximately 10% reach Act 4.
I am currently optimizing an offline "evolving agent" capable of refining deck-building strategies based on data from past runs. Given that an offline simulator provides ground-truth rewards, this appears to be a promising path toward a self-evolving agent.
The Spire agent addresses the challenge of maintaining consistency across long-horizon tasks by delegating deterministic actions to domain-specific tools, while the LLM handles logical tasks requiring reasoning and acts as the "glue" connecting the components.
Currently, about 40% of runs reach Act 3, and approximately 10% reach Act 4.
I am currently optimizing an offline "evolving agent" capable of refining deck-building strategies based on data from past runs. Given that an offline simulator provides ground-truth rewards, this appears to be a promising path toward a self-evolving agent.