# Error recovery

Multi-turn agentic reinforcement learning is widely used to strengthen the long-horizon behavior of large language model agents, yet trajectory-level success rates hide what it does to individual decisions. Across four interactive tasks, DeepCoder-MT, BFCL, InterCode-SQL and FrozenLake, models trained with GRPO are evaluated along two axes, sampling breadth and interaction depth: with more sampled trajectories the RL model consistently beats its base model, but when the turns allowed per trajectory are tightened that edge shrinks, disappears, and on some tasks reverses. Matched-state comparisons show better next-step actions on only 46.8% of DeepCoder-MT states and 42.8% of FrozenLake states, with the gain concentrated in error recovery and later trajectory stages.

Multi-turn agentic reinforcement learning is widely used to strengthen the long-horizon behavior of large language model agents, yet trajectory-level success rates hide what it does to individual decisions. Across four interactive tasks, DeepCoder-MT, BFCL, InterCode-SQL and FrozenLake, models trained with GRPO are evaluated along two axes, sampling breadth and interaction depth: with more sampled trajectories the RL model consistently beats its base model, but when the turns allowed per trajectory are tightened that edge shrinks, disappears, and on some tasks reverses. Matched-state comparisons show better next-step actions on only 46.8% of DeepCoder-MT states and 42.8% of FrozenLake states, with the gain concentrated in error recovery and later trajectory stages.

After a failed tool call, whether an agent retries, fixes the arguments or stops to report depends not only on the error text but on how that feedback enters the conversation. Holding the error content fixed and varying only delivery across 13 open models from 3B to 70B parameters, keeping the failed exchange and appending a runtime note with guidance in a new user turn avoided a severe stopping failure in Llama models; implemented as an opt-in recovery layer inside an inference engine, this form raised ToolMaze success on explicit errors by 14.2 points with 7–8B models and 10.5 points with 32–70B models, with no measurable gain on Loud/Quiet.

After a failed tool call, whether an agent retries, fixes the arguments or stops to report depends not only on the error text but on how that feedback enters the conversation. Holding the error content fixed and varying only delivery across 13 open models from 3B to 70B parameters, keeping the failed exchange and appending a runtime note with guidance in a new user turn avoided a severe stopping failure in Llama models; implemented as an opt-in recovery layer inside an inference engine, this form raised ToolMaze success on explicit errors by 14.2 points with 7–8B models and 10.5 points with 32–70B models, with no measurable gain on Loud/Quiet.