Nature vs Nurture, But for AI Agents

TL;DR

New research from UNC-Chapel Hill shows agents using structured skills outperform agents using raw memory by 15.3% while using 10-20x fewer tokens. The biggest gains come from investing in nurture, not switching models.

Key Takeaways

1. 1.Compressed, structured rules beat raw memory. Not by a little. By a lot.
2. 2.Nature is fixed between model releases. Nurture compounds daily.
3. 3.4 python scripts and one afternoon of work can set up a skill evolution pipeline.
4. 4.After 6 months of nurture, an agent is barely recognizable compared to day one.

nature vs nurture has been debated for centuries. for ai agents, it's not even close. nurture wins.

nature is what the model ships with. raw capability. context window. parameters. the stuff anthropic and openai spend billions training into the weights.

nurture is what you teach it after. structured skills. rules from experience. learned behavior from real mistakes.

nurture wins. and it's not close.

the problem with nature alone

most people run agents on raw capability. they pick the smartest model, give it a system prompt, and hope for the best. when it makes a mistake, they correct it in the chat. next session, the agent forgets and makes the same mistake again.

this is nature without nurture. a gifted kid with no education. all potential, no compounding.

the way most people handle "memory" is stuffing raw conversation history into context. thousands of tokens of "i tried X and it didn't work." it's like studying for an exam by re-reading your entire textbook every time. expensive, noisy, and the agent still misses the lesson buried in the filler.

the research

a paper came out this week from UNC-Chapel Hill called SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning. the researchers built a system where agents learn structured skills from both successes and failures, then recursively evolve those skills over time.

the results: agents using structured skills outperformed agents using raw memory by 15.3%. and they used 10-20x fewer tokens doing it. alphaxiv flagged it as trending and @fly51fly broke down the architecture.

the core insight: compressed, structured rules beat raw memory. not by a little. by a lot. the paper calls it a "SkillBank," a two-tier library of general skills (always loaded) and task-specific skills (loaded on demand).

this mapped perfectly to something boris cherny (creator of claude code) has been saying: tell your agent to update its own instructions after every correction. "claude is eerily good at writing rules for itself." his team does this and says the mistake rate measurably drops.

my version

i've been running a version of this for my own agent (jackson) for a few weeks now. here's the setup:

failure logging. every time the agent gets corrected or something breaks, it logs it:

category: DISCORD
task: format message for discord
what happened: used markdown tables, but discord doesn't render them
root cause: defaulting to familiar formatting instead of platform-specific
fix: use bullet lists for discord, wrap links in <> to suppress embeds
severity: medium

auto-distillation. a weekly script groups failures by pattern and generates rules:

[DISCORD] Default formatting uses tables, but Discord doesn't render markdown tables
Rule: Use bullet lists instead of tables for Discord. Wrap multiple links in <> to suppress embeds
When: Formatting messages for Discord

rule injection. rules get written directly into the agent's instruction files (AGENTS.md). they load automatically every session.

the result. after 2 weeks, i have ~15 evolved rules. the agent stopped making the same discord formatting mistakes. stopped simulating trading data. stopped using the wrong calendar account. stopped using grep when it should use qmd.

these aren't things the model "knows" from training. these are things it learned from getting corrected. that's nurture. and it's exactly what the SkillRL paper predicts: the 200 tokens of structured rules outperform 5,000 tokens of raw conversation history because rules generalize across tasks. raw memory doesn't.

the compounding effect

here's where it gets interesting. nature is fixed. openai ships a new model, you get a step function improvement. then it's flat until the next release.

nurture compounds. every correction adds a rule. every rule prevents a class of failures. after a month you have 30-50 rules. after 6 months, the agent is barely recognizable compared to day one. not because the model changed. because the instructions compounded.

the paper showed this quantitatively: agents with evolved skills kept improving over multiple rounds while agents with raw memory plateaued. the skill library grows but stays compressed. the context cost stays flat while capability keeps climbing.

this is the thing most people miss. they keep switching models looking for better nature. but the biggest gains come from investing in nurture.

how to nurture your agent

you need 4 things:

• a structured failure log (category, root cause, fix, severity)
• a distillation script that groups failures and generates rules
• a way to inject those rules into your agent's context (CLAUDE.md, system prompt, config files)
• a cron job to run it periodically

4 python scripts. one afternoon of work. your agent starts compounding improvements instead of repeating mistakes.

sources:

• SkillRL paper (UNC-Chapel Hill, 2026): the research behind structured skills vs raw memory
• @bcherny on self-updating CLAUDE.md: the practical version from the creator of claude code
• @bcherny's full 10 tips thread: where a lot of my setup originated
• my previous article on 25 things i learned using claude code

everyone is obsessing over which model to use. nature matters. but nurture is where the edge is. and unlike nature, nurture is something you control.

---

related reading

• 25 Things I Learned Using Claude Code Every Day
• How I Set Up 4 AI Agents on One Discord Server
• Your Meeting Notes Are Useless Until You Do This

Frequently Asked Questions

What does nature vs nurture mean for AI agents?

Nature is what the model ships with - raw capability, context window, parameters. Nurture is what you teach it after - structured skills, rules from experience, learned behavior from real mistakes.

What is SkillRL?

SkillRL is a system from UNC-Chapel Hill research where agents learn structured skills from both successes and failures, then recursively evolve those skills over time. Agents using structured skills outperformed raw memory by 15.3%.

How do I nurture my AI agent?

You need 4 things: a structured failure log, a distillation script that generates rules, a way to inject rules into your agent's context, and a cron job to run it periodically.