@TanmayGradient

Creator of TanML | Agentic AI Systems | Senior Quantitative Modeler (AI/ML)

Joined April 2026
  • Tweets58
  • Following65
  • Followers11
10Media
First AIE Code Summit speakers: @poteto, Principal Engineer @ SpaceX @steve_yegge, Engineer, Writer, Advisor @zachbruggeman, Principal Software Engineer @ Ramp Nov 10-12, SF. Come hang out with us. Apply at ai.engineer/code
AI Engineer Code Summit speaker announcement featuring Lauren Tan, Principal Engineer at SpaceX; Steve Yegge, Engineer, Writer, Advisor, Independent; and Zach Bruggeman, Principal Software Engineer at Ramp. Black and white artwork with three speaker portraits and yellow text: November 10–12, SF. Apply at ai.engineer/code.

ALT AI Engineer Code Summit speaker announcement featuring Lauren Tan, Principal Engineer at SpaceX; Steve Yegge, Engineer, Writer, Advisor, Independent; and Zach Bruggeman, Principal Software Engineer at Ramp. Black and white artwork with three speaker portraits and yellow text: November 10–12, SF. Apply at ai.engineer/code.

13
25
9
325
779,751
EvoUndo: Recoverability-Constrained Self-Evolution for LLM Agent Harnesses Tanmay Sah, Dolly Sah, Harshul Jain, Tanya Sah arxiv.org/abs/2608.28363 [𝚌𝚜.𝙰𝙸]
LLM agents increasingly modify their own prompts, tools, middleware, resources, and execution harnesses at runtime. Such self-evolution can improve capability, but a successful mutation may leave persistent effects that cannot be safely reversed in states different from the one in which it was created. We introduce EvoUndo, a framework for representing, synthesizing, diagnosing, and independently verifying recoverability of model-generated self-modifications across counterfactual states. Across 600 unseen one-shot self-evolution tasks, we identify 197 capability-improving mutations that fail recoverability verification. Under the original recovery representation, conventional repair strategies recover 0/197 of these natural failures. Deterministic oracle analysis recovers 48/197 under the original recovery language L0, while the extended recovery calculus increases empirical oracle recovery to 191/197. A protocol-locked 2x2 grounding-by-expressivity intervention then separates two bott

ALT LLM agents increasingly modify their own prompts, tools, middleware, resources, and execution harnesses at runtime. Such self-evolution can improve capability, but a successful mutation may leave persistent effects that cannot be safely reversed in states different from the one in which it was created. We introduce EvoUndo, a framework for representing, synthesizing, diagnosing, and independently verifying recoverability of model-generated self-modifications across counterfactual states. Across 600 unseen one-shot self-evolution tasks, we identify 197 capability-improving mutations that fail recoverability verification. Under the original recovery representation, conventional repair strategies recover 0/197 of these natural failures. Deterministic oracle analysis recovers 48/197 under the original recovery language L0, while the extended recovery calculus increases empirical oracle recovery to 191/197. A protocol-locked 2x2 grounding-by-expressivity intervention then separates two bott

1
1
1
5
459
Tanmay Sah, PhD retweeted
swyx
@swyxOct 16
it finally happened, i'm in a zoom with 3 ai's golden age of solipsism
100
3
6
145
16,769
Tanmay Sah, PhD retweeted
swyx
@swyxOct 14
this is what talent density means at @aidotengineer - @SydneyScouts runs speakers at AIE!!
Alex Volkov
@altryneOct 14
When I tell you @swyx is a theater kid at heart, this is what I mean! Pipes!
30
3
1
70
11,252
Building AI for finance? Bring the problem your team is stuck on. AI Engineer New York is now on @LumaHQ. Oct 12 to 14. Sheraton New York Times Square. Conference tickets required. Luma RSVP does not include admission. Workshops are an add-on. luma.com/6i4mqx8v
🤖 Made with AI
AI Engineer New York 2026 conference cover in charcoal, white and warm gold. The official AI Engineer New York logo sits above NEW YORK 2026 and the words Where AI Engineering Meets Wall Street. A sunset view of the Manhattan skyline and Statue of Liberty fills the lower half. Gold arcs cross the image. Footer, October 12 to 14, NYC.

ALT AI Engineer New York 2026 conference cover in charcoal, white and warm gold. The official AI Engineer New York logo sits above NEW YORK 2026 and the words Where AI Engineering Meets Wall Street. A sunset view of the Manhattan skyline and Statue of Liberty fills the lower half. Gold arcs cross the image. Footer, October 12 to 14, NYC.

1
7
2,106
UndoBench: Separating Task Competence from Recovery Capability in Tool-Using AI Agents Dolly Sah, Tanmay Sah, Harshul Jain, Tanya Sah arxiv.org/abs/2610.05622 [𝚌𝚜.𝚂𝙴 𝚌𝚜.𝙰𝙸] 💬Code: github.com/tradertanmay/undo…
Tool-using AI agents are increasingly deployed across enterprise software systems, yet widely used benchmarks primarily evaluate nominal task completion, conflating baseline planning competence with operational fault recovery. We introduce UndoBench, a benchmark spanning 36 base workflows and 36 fault scenarios across 8 enterprise domains, decoupling task competence from recovery capability via counterfactual paired trials under identical seeds alongside wire-level effect-history and environment-state oracles. On 12 held-out TEST workflows across two open-weight models, two frameworks, and three recovery paradigms (5,760 executions / 2,880 paired trials) in the frozen lost-acknowledgment study, nominal competence reached 83.54% while conditional recovery success rate (CRSR) fell to 46.72%, with naive retry producing duplicate external effects in 53.33% of trials. Extensions to commercial API models reproduced this competence-recovery separation. Evaluations across complementary executi

ALT Tool-using AI agents are increasingly deployed across enterprise software systems, yet widely used benchmarks primarily evaluate nominal task completion, conflating baseline planning competence with operational fault recovery. We introduce UndoBench, a benchmark spanning 36 base workflows and 36 fault scenarios across 8 enterprise domains, decoupling task competence from recovery capability via counterfactual paired trials under identical seeds alongside wire-level effect-history and environment-state oracles. On 12 held-out TEST workflows across two open-weight models, two frameworks, and three recovery paradigms (5,760 executions / 2,880 paired trials) in the frozen lost-acknowledgment study, nominal competence reached 83.54% while conditional recovery success rate (CRSR) fell to 46.72%, with naive retry producing duplicate external effects in 53.33% of trials. Extensions to commercial API models reproduced this competence-recovery separation. Evaluations across complementary executi

1
1
1
70
Tanmay Sah, PhD retweeted
Yingcheng Charles Wu
@Charles_Y_WuOct 5
Excited to introduce JEPA-Anything! One world model that works across molecules, cells, fluids, patients, robots, or even weather. Let's World Model Everything! 📄Paper: arxiv.org/abs/2609.20800 💻Code: github.com/Gen-Verse/JEPA-An…
50
235
20
1,776
131,817
Tanmay Sah, PhD retweeted
swyx
@swyxOct 2
It's time to get serious about Security x AI. One of the joys of doing AIE is helping others start their own high quality conferences in their countries and domains. Every single 2025 partner has come back — and the 2nd AI Security Summit is now way more important than before. Proud to be back as a founding partner with @snyksec! Join the event: aisecuritysummit.com/ The landscape has completely transformed since we started this a year ago. An exploding number of rogue agents, more breaches, more AI-driven attacks. now more than ever, security must be at the center of the AI conversation. With so much more hacks and vulnerabilities, this stuff is no longer theoretical...

AI Security Summit San Francisco 2026

AI Security Summit is coming to San Francisco, CA on October 15, 2026. The flagship event. One full day. One room. No filler.

aisecuritysummit.com
85
8
2
100
13,408
Tanmay Sah, PhD retweeted
swyx
@swyxSep 23
can confirm. ran @latentspacepod AINews side by side with 6 Sol and the difference was night and day: latent.space/p/ainews-claude… 5.5 Opus is the new default model for AINews going forward. so much more concise and tasteful reporting, with much less slopese than even 5 Opus.
Sholto Douglas
@_sholtodouglasSep 22
also important news we fixed the writing
158
14
5
230
42,984
Tanmay Sah, PhD retweeted
Omar Khattab
@lateinteractionAug 11
part of this is in how ~everyone with a platform here speaks about LLMs for the past few months i rely on LLMs as excitedly as anyone, but... are we using the same models, guys?! the frontier is still dumb as a brick 20%+ of the time and needs more hand-holding than a freshman it's just that for some VERY specific types of work, like all the beautiful discovery announcements in the recent past, this hand-holding was painstakingly and expensively done by the labs for us (fantastic btw) but... 99% of the time, i am not searching under the lamppost of the AI lab's over-optimized workflows, so that high cost of alignment to the task is mine to handle and i absolutely do it because it's extremely valuable, and if it needs to be said, i think the progress has been far beyond what i'd have guessed and it's only gonna get far better from here. but don't let the jagged frontier of capabilities delude you into thinking that things are plainly "superhuman" at any *coherently broad* set of capabilities, because A LOT remains to be done (and i think it will be)
Omar Khattab
@lateinteractionAug 8
The vibe and language here have been drifting wholesale away from sensible, nuanced takes on the technical topics that animate me. Been oscillating on whether I should essentially just check out other than amplifying releases, as I've done for the bulk of this year to date. Versus forcing myself to articulate some of this in the long-form style of my 2023-2025 rants. So far, our collective goldfish memory here doesn't help make it feel worthwhile.
30
33
4
351
30,945
Tanmay Sah, PhD retweeted
Lunar
@LunarResearcherAug 3
Andrej Karpathy just explained the 5 shifts turning LLMs into agentic systems. 00:00 - Memory turns chat into personal AI 06:41 - Multimodal AI reads the world 16:58 - Thinking models solve harder tasks 24:51 - Search makes LLMs live 30:58 - Tools turn LLMs into workers Most people are still treating LLMs like chatbots. Karpathy is showing the full stack: Memory → Vision → Reasoning → Search → Tools Prompting is the old workflow. Agentic systems are the new one. This 40-minute talk is worth more than most paid AI agent courses. Bookmark and watch it before everyone catches up. Then read how to turn LLMs into self-improving agent loops below
Lunar
@LunarResearcherJul 28

Agent Harness Engineering vs Loop Engineering vs Graph Engineering

Article · @LunarResearcher
36
466
6
2,780
356,796
Tanmay Sah, PhD retweeted
Vikas gupta
@vicky_grokJul 10
Anthropic pays $750,000+ a year for engineers who can build LLM architectures from scratch. Stanford taught the entire thing in 1 hour lecture & released it for free. Bookmark & watch this today before someone takes it down and read this article "How claude uses context"
Vikas gupta
@vicky_grokJun 19

How Claude Uses Context to Outperform Better Prompts

Article · @vicky_grok
6
25
1
106
25,039
Tanmay Sah, PhD retweeted
Zecheng Zhang
@zechengzhJul 9
Tomorrow I’ll show how to unify your data, services, and systems into a single virtual filesystem for AI agents using Mirage github.com/strukto-ai/mirage Instead of building custom integrations for every tool, expose everything through a familiar filesystem interface. Join if you’re curious about the future of agent system. #AIAgents #AgentSystem #VirtualFilesystem #Mirage

GitHub - strukto-ai/mirage: The World's First Virtual Terminal for AI Agents

The World's First Virtual Terminal for AI Agents. Contribute to strukto-ai/mirage development by creating an account on GitHub.

github.com
cyber•Fund
@cyberfundJul 7
Less preaching, more practice. This Thursday three founders show what their in-production agents can do. An overnight software factory, graph memory, a whole company in one terminal. Live Q&A after every demo.
3
1
1
11
1,044
Tanmay Sah, PhD retweeted
Yoonho Lee
@yoonholeeeJul 7
Thanks for discussing meta-harness @lilianweng ! A very clearly written summary of the state of harness engineering + RSI. I often get the question"what can a harness actually improve?" which i only had partial answers to; this post lays out the whole picture in one place
Lilian Weng
@lilianwengJul 7
new post on harness engineering for AI self-improvement: lilianweng.github.io/posts/2… It is hard to forecast how much the future of RSI will rely on harnesses. Likely harness engineering will evolve in the direction of self-improvement and enable auto-research, and, in turn, smarter models keeps harnesses simple. Even when many harness improvement get eventually internalized into core model, the need to specify goals and context will not disappear.

Harness Engineering for Self-Improvement

The concept of recursive self-improvement (RSI) dates back to I. J. Good (1965), where he defined an “ultraintelligent machine” as a system that can surpass humans in all intellectual activities and...

lilianweng.github.io
3
11
115
10,391
Tanmay Sah, PhD retweeted
Clo Willaerts@bnoxJun 14
TIL verifier tax: an AI agent can complete a task and still fail. It might get the right outcome while breaking a policy, skipping an approval step, or exposing sensitive data. The task is done, but the process wasn't safe. Counts as success or failure? dl.acm.org/doi/full/10.1145/…

The Verifier Tax: Horizon Dependent Safety--Success Tradeoffs in Tool Using LLM Agents | Proceedi...

dl.acm.org
1
1
161
Tanmay Sah, PhD retweeted
Xiangyi Li
@xdotliMay 29
Ty @heathercmiller for organizing the amazing @CAISconf event and @trq212 for the amazing talk! Had hella fun this week 🎉🎉🎉
3
27
1,731
Tanmay Sah, PhD retweeted
Abhay Puri
@AbhayPuri98Jun 1
Really happy to share that our paper -“Malice in Agentland: Down the Rabbit Hole of Backdoors in the AI Supply Chain” has won Best Paper (Outstanding Problem Paper) at @CAISconf 🏆 TL;DR - Poison ~2% of an AI agent’s fine-tuning traces and you can plant a trigger-activated backdoor that leaks confidential data >80% of the time. Guardrails completely miss it. (and thanks GPT for editing me into the team photo so cleanly nobody can tell I wasn’t actually there 😅)
🤖 Made with AI
2
6
1
14
1,868
Tanmay Sah, PhD retweeted
Braden Hancock
@bradenjhancockJun 1
@CAISconf was sold out but small, focused on a particular well-scoped area (compound agentic systems). Some workshop papers were submitted as late as just a couple weeks before (fresh content), and good thing, too! Agentic systems already look quite different today than even 3 months ago. Perhaps the systems focus naturally attracts researchers who are inclined to build prototypes and libraries (not just publish), but I was still pleasantly surprised to see how many members of the Laude community had independently identified this conference as one they wanted to submit to and attend. On top of that, throwing a lounge after hours with food, drinks, demos, salon-style conversation, (and a mini podcast studio, because why not?) made it easy to get an even higher concentration of interesting people with interesting half-baked ideas in a particular niche that happens to be on fire right now. Loved the experience. Well-done, CAIS organizers (@heathercmiller, @lateinteraction, @matei_zaharia, @deeptir18, et al). Same time, same place next year?
1
3
2
15
4,661
One problem, two solutions, one through-line: agentic systems need real guarantees, not heuristics. All three, plus 60 more papers: caisconf.org/program/2026/pa…
1
3
323
Tanmay Sah, PhD retweeted
Atal
@ZabihullahAtalApr 18
Stanford just released a 1.5-hour lecture on “LLM Architecture.” This is the exact thing systems engineers at Anthropic and OpenAI require to understand at a deep level. Give it some time. This might be the highest-ROI learning you do this month.
34
670
15
4,607
472,877
Load more