Rohan Kumar

research @greptile, quantum computing phd student @Yale

[Rohan Kumar @rohankumar]

@_WEEXIAO great stuff guys!

[roon @tszzl]

the greatest whitepill of all is that local minima are rare in high dimensional spaces

[Daksh Gupta @dakshgup]

my sister is interviewing at startups after i convinced her to. i met with her today, and she remarked how much more excited and happy the people she met in startup world were than the ones she had met in banking/consulting.

both work many hours, but startup employees share this fact with glee rather than dismay.

this is of course a generalization, but it made me think of something @hungrysohan said once -

“how fortunate we are to be tired in pursuit of a challenge of our choosing”

[Rohan Kumar @rohankumar]

@dakshgup @soohoonchoi @vaishaaant shared uber to ipo bell baby

[Rohan Kumar @rohankumar]

opus 4.8 has made frontier coding multipolar again

[Nick Khami @skeptrune]

i think about code review a lot

it's really broken now. absolutely impossible to review all the PRs manually

for a long time cursor bugbot felt like the best solution. today it's without a doubt @greptile

feels like a codex vs. claude code vibe shift. good reminder to always keep building

[Rohan Kumar @rohankumar]

me when I gradient descent my agent md to get my agent to simulate an agent on top of my agent so it knows how to be a better agent than my agent

[Muratcan Koylan @koylanai]

Gradient descent for SKILL.md files sounds interesting, maybe a bit complex but it's becoming a real part of agent harness.

SkillOpt is one of the first papers to treat markdown skill files as trainable parameters and provides a proper optimization framework for them.

A few things I learned that you should consider too.

1. The validation gate is the only thing that matters in a self-editing loop.

Held-out set, strict improvement, ties rejected. End-to-end, their best skills land with 1 to 4 accepted edits total. If your "self-improving agent" is accepting most of what it proposes, you're shipping slop.

2. Bounded edits are better than full rewrites. 4 to 8 edits per step is the sweet spot.

Remove the budget and performance collapses. This is the textual analog of learning rate, and it transfers to any LLM-as-author loop. If you're using an agent to refactor your docs, your prompts, or your skills, cap the diff size.

3. Compactness wins. Median final skill: ~920 tokens.

Skills do not need to be long. They need to be high-signal. Most skill files I see are bloated because length feels like effort. It isn't.

4. The harness is becoming less important; the skill is becoming more important.

A Codex-trained skill ported into Claude Code hit +59.7 points on SpreadsheetBench. Procedural knowledge is more general than the runtime that
produced it.

5. Frozen model + trained context is the practical adaptation.

GPT-5.4-nano with a SkillOpt'd skill ≈ frontier behavior on procedural benchmarks. Cheaper, portable, inspectable, zero inference-time cost. This is
the answer to "how do we adapt a frontier model for our domain" for almost everyone who isn't training their own models.

6. Verification is the bottleneck.

Every gate in this paper depends on an auto-grader. That works for benchmarks. It fails for writing, design, and strategy, exactly the open-ended work we want to automate. Whoever builds the verifier for open-ended tasks owns the next stage.

There are also two leassons I learned while shipping v2.3.0 of my Context Engineering Agent Skills repo, measured across composer-2, claude-opus-4-7,
gpt-5.5, and gemini-3.1-pro via the @cursor_ai SDK:
- Description and body are two different surfaces. The router only sees the description. The agent sees the body once activated. They can quietly disagree, and only end-to-end task tests catch it.
- Aggregate accuracy is the wrong unit. When I rewrote three descriptions, the corpus average moved ~1pp. Individual skills moved 23–25pp. Per-skill effect size is where the action is.

Also, in Feb 2026 I shared a piece called Personal Brain OS arguing that the markdown file is a first-class substrate for agent state. SkillOpt is the optimizer-shaped version of that same argument: not "store memory in files" but "treat files as trainable parameters with proper optimization machinery around them." That's the move from static to measured.

The fast/slow split they describe already lives implicitly in the digital-brain-skill repo:
- voice-guide and tone-of-voice.md are slow-state (rarely touched)
- posts.jsonl and bookmarks.jsonl are fast-state

What SkillOpt adds that I didn't have is a protected section invariant, a structural guarantee that fast edits cannot overwrite slow lessons. Removing that mechanism cost them 22 points on SpreadsheetBench. Worth borrowing.

If you're building agents, SkillOpt: Executive Strategy for Self-Evolving Agent Skills is a good paper to read: https://t.co/ZS9SZXQ6Mv

[arman @ksw_arman]

it's crazy how @greptile has had such a noticeable improvement in the last few months. i've never seen an agent at that scale improve drastically so fast

[Rohan Kumar @rohankumar]

@BenBackus4

[Rohan Kumar @rohankumar]

@christineehan fact check: there was only one Korean in this photo

[Rohan Kumar @rohankumar]

@BenBackus4 how does he keep pumping out bangers

[Rohan Kumar @rohankumar]

why do so many people suddenly have performative subscriptions to Nature

[Rohan Kumar @rohankumar]

@brianchew @aiDotEngineer @vaishaaant @dakshgup @ExaAILabs @yangshun @mint0themax @adlinzainal @swyx @dakshgup does this mean I get to buy the b200s now

[Rohan Kumar @rohankumar]

@prenkist :stagingisdown:

[Daksh Gupta @dakshgup]

launching 5 things:

1. multi-repo context support

2. rebuilt web app for super large orgs

3. integrations with claude/codex/devin

4. .greptile/rules files

5. rebuilt learning so greptile maintains internal docs about your company

[Rohan Kumar @rohankumar]

@BenBackus4 Boutta procrastinate so hard with this

[Rohan Kumar @rohankumar]

@greptile nah i trust greptile with my life rn

[Daksh Gupta @dakshgup]

more than a quarter of all code reviewed by greptile is now written by "background agents" - completely autonomous e2e coding agents like devin.

we analyzed millions of PRs written by background agents and compared their
- code churn rates
- revert rates
- bugs per PR
against human baseline.

read the full post on our blog!

[Rohan Kumar @rohankumar]

@BenBackus4 @vboykis You can use it whenever you want if you find the Easter egg on the @greptile dashboard :))

[Rohan Kumar @rohankumar]

@vimberb73649 @thsottiaux seconded ^^