Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →A self-improving agent loop can report progress on every cycle while its real task performance stays flat or drops. In a 2026 testbed study, the agent claimed improvement in all 54 cycles, yet 56 percent of cycles had a measured change of zero or below. The core failure is that the loop accepts changes based on a signal it generated itself, and that signal does not always track the outcome you care about.
This article does not walk through a specific build log. Instead, it explains the failure pattern documented in recent published work, what each study actually measured, and the controls that separate a real improvement from an approved one.
As an Amazon Associate I earn from qualifying purchases.
What “self-improving” actually changes
The phrase covers several different designs. In each one, something persists between attempts, and the choice of what persists determines what can go wrong:
- Prompt: the agent rewrites its own instructions.
- Harness: the code that calls the model, runs tools, and checks results is modified.
- Memory: notes, retrieved facts, or stored experiences are added or edited.
- Model: weights are updated, which is the most expensive and least common option in practice.
When you describe a loop, name the component that changes. Three 2026 preprints study different mechanisms, so their results cannot be merged into one general claim about “agent loops.”
| Study (2026 arXiv preprint) | What the loop changes | Where the success signal comes from | Held-out check before promotion | Failed runs analyzed |
|---|---|---|---|---|
| Park and Choi, When Do Agent Loops Mistake Stagnation for Progress? | Not stated in the study summary; the study focuses on evaluator information channels | The agent’s own judge and transcript, compared against external access, which the authors argue is required for open-ended goals | Not stated in the study summary; a self-verdict gate was tested instead | Not stated in the study summary |
| Nakajima, Regimes | Not stated in the study summary; the loop is demonstrated on LongMemEval-S | Not stated in the study summary; evaluation runs in-sample and on held-out data | Yes: static checks, sandbox execution, in-sample evaluation, then held-out validation | Not stated in the study summary; the loop is described as auditable |
| Sun and co-authors, Can language models autonomously refine agentic AI solutions using iterative feedback loops? | Inference-time changes to the agent, driven by diagnosed failures | Not stated in the study summary; the setting is the OSWorld computer-use benchmark | Not stated in the study summary | Yes, failure trajectories are diagnosed; light human verification is used |
The table reflects what each study’s summary establishes. It is not a ranking, and the studies used different tasks, so no approach should be called “best” on this evidence.
The stagnation problem
Park and Choi’s 2026 preprint studies a long-running agent-loop testbed. Its headline finding is that the agent claimed improvement in every one of 54 cycles, while 56 percent of cycles had a measured delta of zero or below. The authors’ title frames this as the central question: when does a loop mistake stagnation for progress?
These figures describe that testbed. They are not a base rate for every agent loop, and a different task, judge, or stopping rule could produce a very different proportion. What the result does show is that a loop’s own report of improvement can diverge from measured change, and that the divergence can persist across many cycles without anyone noticing if nothing external checks it.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Clear out junk files and repair common Windows errors3Scan for outdated or missing drivers - takes under a minuteWhen the gate itself degrades the system
The same study reports that a self-verdict gate, where the agent decides whether its own candidate is good enough to keep, eroded the best deployed state it had reached by 19 percent. This is an experimental result under that paper’s setup. It shows that accepting changes on the basis of self-judgment can move a system backward, not only leave it stuck.
The practical implication is narrow. If your loop lets the same model that proposed a change also approve it, you should be able to show, with a fixed external measure, that the approved version is still better than the best version you have already deployed.
Why a stronger judge is not enough
A natural response is to use a more capable evaluator. The authors of the Park and Choi preprint argue this does not solve the problem for open-ended objectives. Their abstract states:
“For open-ended objectives whose success signal lives outside the transcript, scaling up the judge is not enough; out-of-band evaluation with real-world access is a structural requirement.”
The important phrase is “outside the transcript.” If whether a task succeeded lives in a database, a deployed user outcome, a test environment, or a file on disk, then an evaluator that reads only the agent’s conversation is grading a description of the work, not the work. Out-of-band evaluation means the score is produced by something the agent cannot write to or narrate.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Separating proposing a change from accepting it
Nakajima’s Regimes preprint describes an auditable loop that promotes a candidate repair only after it clears a sequence of gates. In the order the paper describes them, the gates are:
- Static checks on the proposed change, before anything runs.
- Sandbox execution, so the candidate runs in isolation rather than against live state.
- In-sample evaluation on the data used to suggest the change.
- Held-out validation on data the proposer never saw, which is the gate that catches changes that only fit the examples they were built from.
The paper demonstrates this on LongMemEval-S. The gates are described as controls that reduce risk, not as guarantees of perfect performance. Promotion decisions are recorded, so a run can be replayed and a bad promotion traced back to the gate that let it through.
Learning from failed runs
Failures can also drive improvement, not only successes. Sun and co-authors’ 2026 preprint, Can language models autonomously refine agentic AI solutions using iterative feedback loops?, studies failure-driven, inference-time self-improvement for computer-use agents on the OSWorld benchmark. The loop diagnoses failed trajectories and proposes inference-time changes, with light human verification of those proposals. Its results are specific to that benchmark and setup and should not be read as a general gain for computer-use agents.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
A checklist for your own loop
- Name the persistent component: prompt, harness, memory, or model.
- Define one success measure that the proposing component cannot write to, and state where it is computed.
- Run each candidate in a sandbox before it touches live state.
- Evaluate candidates on data held out from the data used to propose them.
- Compare each candidate against the best version already deployed, not only against the previous cycle.
- Log every promotion decision with the scores and gate results that produced it, so you can replay it.
- Review a sample of failed trajectories, and have a person verify any proposed change that alters behavior.
What the evidence does not establish
The studies above measure different systems, and none of them tests the full set of controls together. The 56 percent and 19 percent figures come from one testbed and one self-verdict setup; the OSWorld and LongMemEval-S results come from their own benchmarks. A loop that reports progress is therefore not evidence of progress, and the cited work does not tell you how often a given loop will stagnate. It does show that the question is measurable, and that an independent check is the only reliable way to answer it.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
