Co-founder at SquareDiff
Reach me at: mayank@squarediff.com
axiomatic truths
- All agents are composed of feeding tokens into an LLM in a loop(s)
- An agent is only as good as its model and the relevance of the tokens you pass in at each turn
thesis
- For a given task, there is some precise combination of tokens to the model that maximizes answer quality
- Each step of every task has some tokens_in array that leads to the highest possible output quality
- The harness can be autonomously iterated on in parallel to hillclimb towards this ideal input
what I'm building
SquareDiff allows you to:
- Deploy experiments on your agent's harness in parallel
- Allow agents to autonomously propose changes to your harness
how it started
At Meta I started to notice issues with the traditional agent building process:
- Engineers observe an error
- Read the trace to find root cause
- Update system prompt to fix trajectory
- Incur prompt debt, token and accuracy waste, fail to consider other harness improvement options, etc.
Halfway into the agent's lifecycle, we realized the system prompt was consuming 70k tokens of the 250k window. (Throwing away 20-30% of our accuracy at the first query)
There were two issues:
- There's multiple potential non-obvious approaches to fix an error with your agent: system/tool prompting, automatically feeding information, delegating to expert subagents, skills, etc.
- Agents can reason about how to build agents faster and better than humans can
work
SquareDiff- Data Science Agent with 10k DAU
- Enabled the agent's first context window management system
- Worked on a lot of the early fun problems around custom tooling for agents
- How do you intelligently feed tabular data into an LLM designed to handle natural language?
- i.e. "Our agent keeps writing SQL queries without knowing the shape of the data, do we prompt it to call the same data tool, force it to call the tool, or connect the sample data to…etc."
- Experimental projects with LLMs
- Enabling natural language querying of SQL databases
- Maximizing CPU/GPU utilization
- Infra monitoring for cloud oversight
projects
- Lorn AIEnable ecom store owners to sell their products on ChatGPT/ACP providers
- Visa Social ReviewIn early 2026 the US government began requiring h1b applicants to make their social media public to be reviewed. Browser use agent to read all your posts and flag potentially risky content you may have forgotten
- CravrScraped every restaurant menu in SF to make an agentically queryable database of food options, like "low calorie high protein brunch" or "chicken caesar salad"
- MisfareOpen source env variable diff tool
- Let Me Check FirstAssign a hotkey to allow local hosted LLMs to rewrite your AI queries with better formatting and spellcheck
writing
interests
- SpiritualityI spent a huge portion of my teenage years thinking I was going to skip college to move to the mountains and be a monk. Turns out, building a company has taken priority temporarily. Read Autobiography of a Yogi.
- Stand Up ComedyA habit I started in college and haven't picked up again since Cursor came out. Fewer things will shatter your ego like trying to make a crowd laugh while they just stare back at you :) Learned a lot here
- PokerBy far the best way to make friends in SF
- Chessplayed competitively for 6 years as a kid and lost my appetite for the game. All games played are while waiting for coding agents to finish
A merchant sent his son to learn the Secret of Happiness from the wisest of men. The young man wandered through the desert for forty days until he reached a beautiful castle at the top of a mountain. There lived the sage that the young man was looking for.
However, instead of finding a holy man, our hero entered a room and saw a great deal of activity; merchants coming and going, people chatting in the corners, a small orchestra playing sweet melodies, and there was a table laden with the most delectable dishes of that part of the world.
The wise man talked to everybody, and the young man had to wait for two hours until it was time for his audience.
With considerable patience, he listened attentively to the reason for the boy's visit, but told him that at that moment he did not have the time to explain to him the Secret of Happiness.
He suggested that the young man take a stroll around his palace and come back in two hours' time.
"However, I want to ask you a favor," he added, handing the boy a teaspoon, in which he poured two drops of oil. "While you walk, carry this spoon and don't let the oil spill."
The young man began to climb up and down the palace staircases, always keeping his eyes fixed on the spoon. At the end of two hours he returned to the presence of the wise man.
"So," asked the sage, "did you see the Persian tapestries hanging in my dining room? Did you see the garden that the Master of Gardeners took ten years to create? Did you notice the beautiful parchments in my library?"
Embarrassed, the young man confessed that he had seen nothing. His only concern was not to spill the drops of oil that the wise man had entrusted to him.
"So, go back and see the wonders of my world," said the wise man. "You can't trust a man if you don't know his house."
Now more at ease, the young man took the spoon and strolled again through the palace, this time paying attention to all the works of art that hung from the ceiling and walls. He saw the gardens, the mountains all around the palace, the delicacy of the flowers, the taste with which each work of art was placed in its niche. Returning to the sage, he reported in detail all that he had seen.
"But where are the two drops of oil that I entrusted to you?" asked the sage.
Looking down at the spoon, the young man realized that he had spilled the oil.
"Well, that is the only advice I have to give you," said the sage of sages. "The Secret of Happiness lies in looking at all the wonders of the world and never forgetting the two drops of oil in the spoon."