Current AI agents often remain static once deployed, failing to adapt to shifting user needs without costly retraining downtime. MetaClaw addresses this "stagnation gap" by introducing a dual-track learning system that enables agents to grow "in the wild" during actual production use. This framework allows software assistants to refine their own capabilities based on real-world interactions rather than waiting for manual developer updates.

The first track involves skill-driven fast adaptation. When an agent fails a task, an internal "evolver" analyzes the error and synthesizes a new reusable skill. This provides immediate performance boosts without modifying the underlying model weights, allowing the agent to handle 20+ messaging channels with increasing precision.

The second track, opportunistic policy optimization, handles deeper structural changes. By monitoring system inactivity and calendar data via an internal scheduler, MetaClaw triggers cloud-based fine-tuning using LoRA and Reinforcement Learning with a Process Reward Model (RL-PRM). This ensures the agent updates its core logic only when the user is away, effectively "sleeping" to process the day's lessons.

In testing, this autonomous evolution pushed the Kimi-K2.5 model's accuracy from 21.4% to over 40%, nearing the performance of GPT-5.2. By separating support and query data through a versioning mechanism, the system prevents "data contamination," ensuring the agent learns from genuine experiences rather than just memorizing past failures.

Current AI agents often remain static once deployed, failing to adapt to shifting user needs without costly retraining downtime. MetaClaw addresses this "stagnation gap" by introducing a dual-track learning system that enables agents to grow "in the wild" during actual production use. This framework allows software assistants to refine their own capabilities based on real-world interactions rather than waiting for manual developer updates.

The first track involves skill-driven fast adaptation. When an agent fails a task, an internal "evolver" analyzes the error and synthesizes a new reusable skill. This provides immediate performance boosts without modifying the underlying model weights, allowing the agent to handle 20+ messaging channels with increasing precision.

The second track, opportunistic policy optimization, handles deeper structural changes. By monitoring system inactivity and calendar data via an internal scheduler, MetaClaw triggers cloud-based fine-tuning using LoRA and Reinforcement Learning with a Process Reward Model (RL-PRM). This ensures the agent updates its core logic only when the user is away, effectively "sleeping" to process the day's lessons.

In testing, this autonomous evolution pushed the Kimi-K2.5 model's accuracy from 21.4% to over 40%, nearing the performance of GPT-5.2. By separating support and query data through a versioning mechanism, the system prevents "data contamination," ensuring the agent learns from genuine experiences rather than just memorizing past failures.