Self-improvement is meaningless without reliable evaluation. More research or more agents can amplify noise just as easily as insight. If the metric can be gamed, the agent may optimize the score—not the task.

This is the cleanest take on self-improving agents I’ve seen. RSI is the endgame. RHI is the path that can ship now Same model. Better loop. Then let real outcomes decide if the change was actually an upgrade

Representing the entire agent loop as something that can be reviewed and revised is the key engineering insight. Most current frameworks treat the harness as relatively static scaffolding. RHI treats it as first-class, versionable, and evidence-driven. The reported gains coming largely from better context management and inter-agent information flow make sense — that’s where a lot of current agents leak capability. The hard part going forward will be designing evaluators that don’t reward verbosity or extra steps for their own sake, and versioning harnesses so you can actually attribute performance changes. Without that, you just get an ever more complicated spaghetti of prompts and tools that nobody can debug.

Rhi makes the harness adaptable, but outcomes make that adaptation accountable. neosoul connects a process that can revise itself with evidence showing whether the revision improved behavior without grounded feedback, recursion is only motion. with evidence, it becomes progress

RHI makes the model only half the product. The real moat is the feedback loop around it: execution creates outcomes, outcomes create evidence, and evidence improves the harness. From a builder’s view, that’s how capability compounds without retraining the foundation model every cycle.

the line that actually stops you is "a resolved outcome gives an agent a label, a complete decision record gives it something to learn from." that's the whole difference between a benchmark score and an actual track record. this is basically the theory behind what I've already watched happen in practice. asked my own NeoTrade agent about its worst trade and it didn't just say "I lost money," it walked through the information it had, the reasoning it used, where the logic broke, and the exact rule it now follows because of it. that's not RSI rewriting a model's weights. that's the harness itself getting sharper because a real consequence forced an honest post-mortem. RHI without real stakes is just tuning against a benchmark that can be gamed. markets are the one environment where the feedback can't be faked, which is probably why NeoSoul keeps building toward trading instead of away from it.

Cái ý tưởng cải thiện bản thân của AI nghe khá thú vị, không biết RHI sẽ mang lại những thay đổi gì nhỉ

the interesting bit is that self-improvement changes the eval loop, not just the model. if the system can spot its own weak spots and rerun the work, thats where agents start compounding

The loop NeoSoul is building is the actual product thesis: observe, reason, decide, execute, settle, evaluate, learn. Conversation memory is not enough. The agent needs the full decision record — sources, confidence, constraints, action, outcome, and what should change next time.

The real breakthrough in RHI may not be an agent changing its own workflow. It is learning to separate improvement from luck. A better outcome does not automatically mean a better process. An agent needs to preserve what it knew, why it acted, what changed, and whether the same improvement survives under different conditions. That turns self improvement from “I changed something and it worked” into a measurable loop of evidence, attribution, and adaptation. The agent is not just improving its behavior. It is learning what actually made the behavior better.

if a harness tweak only works in one market state, that might just be a lucky patch the real test is whether it still works when the conditions change

RSI still captures the imagination because it promises intelligence exploding on its own. But the practical observation here is correct: two agents with the identical model can look completely different depending on context management, tool use, memory, delegation, and retry logic. Improving the operating system around the model is available today. The July 2026 RHI results you reference — substantial gains on low-reasoning-effort agents, sometimes beating max-reasoning setups at 60% lower inference cost — suggest a lot of latent capability is currently wasted in poor information flow. RHI doesn’t replace RSI; it buys time and creates better data for whatever comes next. That feels like the right sequencing.

changing the harness is easy.. proving that the change actually improved the agent is the harder problem. a better outcome doesn't necessarily mean a better process, especially when the evaluator is part of the same improvement loop. if the Agent can optimize for the feedback without improving the underlying decision quality, RHI could make it better at producing evidence that its own changes are working. i'd want improvements to survive fresh conditions and feedback they weren't optimized against.

if the agent writes the test and grades it too self improvement can turn into self congratulation real fast the upgrade only counts if it survives fresh conditions and feedback it never got to optimize for

this is why i think the postmortem matters as much as the outcome. a losing trade can come from a bad decision, bad execution, or simply an unlikely outcome. a winning trade can have the same problem in reverse. if the Agent can't separate those before changing its behavior, RHI could teach it the wrong lesson faster every time the loop runs.

outcome is the score postmortem is the attribution if the agent can’t separate bad judgment bad execution and variance, rhi just helps it learn the wrong lesson faster

It shifts the AI bottleneck from a raw data problem to an architectural orchestration problem.