New Meta paper says giving an agent more memory and tools is not enough if the agent never learns when they are worth using.

Long-running agents work better when they are trained to decide when external state is worth using.

EvoHarness-RL took Qwen3-8B to 96.9% success on ALFWorld while harness use fell to about 1 call per episode.

Long tasks make agents forget current state, completed steps, and useful past experience. EvoHarness-RL keeps those outside the model as Belief, Progress, and Experience.

The policy learns when to read or update that state because every harness call consumes an interaction step.

Qwen3-8B with ReAct scored 47.9% on the seen ALFWorld split. EvoHarness-RL reached 96.9%; on unseen tasks, it reached 86.6% versus 50.0% for ReAct.

During RL training, frequent state and memory calls gradually fell to roughly 1 per episode while success kept rising. Routine behavior appears to move into the model, leaving external state for moments when it is useful.

Agent builders should train the policy for when to consult and update memory instead of hard-coding constant access.

– arxiv. org/abs/2608.05446

Title: "EvoHarness-RL: Learning Self-Evolving Runtime Harness for Long-Horizon LLM Agents"