An AI agent's task-tracking setup boosted accuracy by 29 points — but which of its parts actually did the work?

This explores which individual pieces of StateM's state management account for most of its accuracy gains. The corpus can't answer that directly. None of the retrieved notes report an ablation that removes StateM's components one at a time. What it does have is one large whole-system result, one design choice singled out as the key move, and a few neighboring findings that suggest where the gains might come from.

The headline number comes from a note on task state management. Separating state from execution, and checking it with independent environment audits instead of trusting what the executor claims, lifted Qwen 3.7-Plus from 51.8% to 80.7% on WeaveBench, a 29-point jump (Can task state management alone improve long-horizon agent performance?). That note doesn't name StateM, so the link is likely but unconfirmed. The comparison is the full system against a baseline. It can't say how much of the 29 points comes from moving state out of the model and how much from auditing it.

The note that does name StateM reports gains on Terminal-Bench 2.1 across several models with the weights frozen: 95.3% on GPT-5.6 and +5.4 points on DeepSeek-V4 Flash. The same runbook carried over to newer models without changes (Can execution harnesses lift model performance without retuning weights?). Gains are reported per model and per benchmark, never per component. That makes the results hard to compare with the WeaveBench jump, and impossible to attribute to a specific piece.

Two neighboring findings suggest candidate drivers, and they are hypotheses only. In one, a stronger model built inference-time harnesses that nearly doubled a weaker model's Theory-of-Mind scores. The gain came mostly from moving unstable reasoning into deterministic code and task-specific routing, not from encouraging longer reasoning (Can a stronger model lift a weaker one at test time without retraining?). LLM Programs work the same way: an explicit algorithm owns the control flow and state, and each model call sees only the context relevant to its step (Can algorithms control LLM reasoning better than LLMs alone?). If StateM follows this pattern, the biggest contributors would be deterministic state handling, hiding irrelevant context, and independent verification.

The gap isn't unique to StateM. Prime Agent organizes state in four levels (weights, context, a persistent REPL with subagents, and disk-backed history) and reports gains on ARC-AGI-3. Its own note says the specific components remain unablated (Can external state caches let models solve harder problems?). Harness papers in this area tend to report that the whole package works without saying which part does the work. To settle this question you would need the ablation table from the StateM paper itself, which this library doesn't yet cover.