AI agents on long optimization jobs succeed less by starting smart than by not quitting — why does persistence win?
Why does persistence in the feedback loop predict agent success better than initial solution quality?
This explores why, on very long optimization tasks, the number of productive try-measure-revise cycles an agent completes tells you more about its success than the quality of its first attempt.
This explores why, on very long optimization tasks, the number of productive try-measure-revise cycles an agent completes tells you more about its success than the quality of its first attempt. In the study behind the question, 17 frontier models worked on 36 expert-built optimization tasks. Most stopped early or burned their time budget without making progress, and the models that kept running benchmark-edit-incorporate cycles did best (Claude Opus 4.6 stood out) What predicts success in ultra-long-horizon agent tasks?. That study shows the pattern but not its cause. Neighboring notes in the corpus point to three likely reasons.
First, a first attempt can only contain what the model already knows, and every loop brings in something new from outside. Pure self-improvement stalls because a model can't reliably verify its own work. The methods that do work bring in an outside anchor such as past model versions, third-party judges, user corrections, or tool feedback Can models reliably improve themselves without external feedback?. A benchmark run is one of those anchors. It also carries more than a score. Feedback says how well an action did and also which way to change it Can scalar rewards capture all the information in agent feedback?. An agent that quits early leaves all of that uncollected, and a brilliant first draft has only used the anchor once.
Second, loops compound while a first draft is a single draw. Reflexion shows that clear success-or-failure signals let an agent write useful self-diagnoses and improve from one episode to the next with no weight updates Can agents learn from failure without updating their weights?. ReasoningBank finds that memory and extra compute multiply each other, so accuracy keeps rising with accumulated interaction history Can agents learn better from their failures than successes?. Under that logic, a mediocre start with many cycles beats a strong start with few.
The finding also says the persistence has to be productive. Burning budget without progress counted as failing, which means the loop has to turn each failure into something usable. SkillRL keeps successes as concrete demonstrations and failures as abstracted lessons, and this beats treating them the same Should successful and failed episodes be processed differently?. Reliable agents also move state and memory into the surrounding harness so a long run doesn't lose track of what it has tried Where does agent reliability actually come from?. An agent that keeps looping without absorbing what it learned only looks persistent.
Third, persistence is a property of the whole run, and first-solution quality is a snapshot of one moment. Agent evaluation is shifting from final answers to whole trajectories, scoring process quality and recoverability as well as correctness How should we evaluate agent behavior beyond final answers?. On long tasks, what an agent does after its first failure separates models better than how good its first answer was. The practical upshot is that long-horizon skill depends partly on stubbornness plus bookkeeping, not only on raw intelligence. The corpus has one study showing the pattern, so the mechanisms above are inferred from related work, not tested directly.
Sources 8 notes
Across 17 frontier models on 36 expert-curated optimization tasks, repeated benchmark-edit-incorporate cycles within a wall-clock budget proved the dominant success predictor. Most models terminated early or burned budget unproductively; Claude Opus 4.6 stood out as persistent.
Pure self-improvement stalls due to the generation-verification gap, diversity collapse, and reward hacking. Reliable improvement methods succeed by smuggling in external anchors: past model versions, third-party judges, user corrections, or tool feedback.
Natural feedback carries two orthogonal types of information: evaluative (how well an action performed) and directive (how it should change). Scalar rewards capture evaluation but discard directional specifics that token-level distillation can recover, making the two complementary rather than redundant.
Reflexion demonstrates that unambiguous environmental feedback (success/failure) enables agents to write useful self-diagnoses and improve across episodes without parameter updates. The binary signal prevents rationalization, and keeping reflections uncompressed preserves their usability.
ReasoningBank shows that storing strategy-level reasoning hints from both self-judged successes and failures outperforms success-only memory and raw trajectory storage. Coupled with test-time scaling, memory and compute compound rather than substitute, creating a novel scaling law where accuracy improves through cumulative interaction history.
Show all 8 sources
SkillRL demonstrates that treating successful episodes as concrete demonstrations and failures as abstracted lessons achieves state-of-the-art performance on complex tasks while using substantially less context than uniform approaches. The asymmetry mirrors human expert reasoning and avoids the degradation seen in uniform consolidation methods.
Research shows reliable LLM agents externalize three cognitive burdens—memory (state persistence), skills (procedural components), and protocols (structured interaction)—into a harness layer rather than relying on model scale alone. The harness unifies these externalities and eliminates the need for the model to solve the same problems repeatedly.
Evaluation of agentic systems shifts evidence from final responses to full interaction sequences, and scoring procedure from correctness alone to process quality, recoverability, coordination, and robustness. This pattern appears across multiple agent benchmarks as a coherent design move.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Useful Memories Become Faulty When Continuously Updated by LLMs
- AgentFly: Fine-tuning LLM Agents without Fine-tuning LLMs
- RLVMR: Reinforcement Learning with Verifiable Meta-Reasoning Rewards for Robust Long-Horizon Agents
- ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory
- Beyond Final Scores: A Systematic Evaluation of Agents for Long-Horizon AI Research and Development
- RRSI: Regularized Recursive Self-Improvement of Agent Harnesses
- SkillClaw: Let Skills Evolve Collectively with Agentic Evolver
- Demystifying Agent Skills: Why They Work-Until They Don't