Adhyyan Sekhsaria (@Adhyyan) profile picture

Founding Engineer

San Francisco
Joined January 2015
Photo by Adhyyan Sekhsaria (@Adhyyan)Photo by Adhyyan Sekhsaria (@Adhyyan)Photo by Adhyyan Sekhsaria (@Adhyyan)
My dream is to reach 100 PRs a day consistently. I was close to 10 PRs a day on a good day of work. With bots its easy to hit 30 PRs a day without being as tired as before!
Meet Devin Bots. From writing firmware to merging thousands of PRs a week, a Devin Bot leads a team of agents to get the job done.
6
3
48
1,368
Adhyyan Sekhsaria retweeted
Since joining @cognition, the @dioxuslabs team has shipped: - macOS VMs - memory snapshot VMs - browser previews + Devin Apps - advanced hybrid computer use - slashed VM costs by over 50% - improved devin startup by 90% - ssh to devin - 60fps quic vnc - embedded editor with lsp
25
11
3
264
15,863
Adhyyan Sekhsaria retweeted
we ordered these 2 weeks ago
58
8
2
470
35,090
Adhyyan Sekhsaria retweeted
Best AI product of 2026 for me there are 4 - @DevinAI, @eve, @slashyemail and @greptile > devin set the bar for what a coding harness looks like. I can actually gauge how much people know about harnesses from their opinion on Devin. > eve formalized the idea of serverless agents. it’s an incredibly well written framework with solid fundamentals and good surrounding infrastructure. > because of slashy I don’t look at my email anymore. and it’s not because their UI sucks, it’s because they have this incredible iMessage agent that just handles my inbox and scheduling. > I can’t really imagine writing software without greptile. you can distinguish codebases I’m serious about from whether I use greptile or not
Best AI product of 2026?
9
7
1
63
5,688
Adhyyan Sekhsaria retweeted
Scott Wu vs. the models
this is my quant
93
274
101
5,031
516,578
Adhyyan Sekhsaria retweeted
Feels good to finally be able to talk about what you spent the last half of the year building
Introducing Listen Twins. 1:1 digital twins of real people, each built from an in-depth interview. Ask them anything and 100+ twins answer in seconds.
4
2
1
18
831
monosemantic neurons be like
19
76
6
1,415
30,007
Adhyyan Sekhsaria retweeted
Devin is rapidly taking over Codex and Claude Code internally at ploy we're essentially doing 90% of dev work out of slack what the heck is going on
36
14
10
256
40,213
Adhyyan Sekhsaria retweeted
wow… devin is actually great now wtf. the ads were spot on
8
14
1
147
14,585
Adhyyan Sekhsaria retweeted
cool to see cognition doing "dreaming" - memory needs an offline cleanup loop, not just better retrieval, otherwise stale records pile up open q is how inferred memories get validated before the agent uses them glad it's an open standard too - memory should be open teapawt.net/cognition/status/21071…
Introducing Dreaming: Across sessions, Devin builds a memory graph of how you like to work At night Devin self-improves its memory to remove stale records and discover latent information We are creating an OSS standard called Agent Memory Repo: cognition.com/agent-memory-r…
71
11
4
159
17,923
Adhyyan Sekhsaria retweeted
too long didn't read
10
2
2
358
19,382
A git repo completely operated by the agent is much more powerful than a single file on your local machine. It handles merge conflicts, observability and scales to much more information. Every cloud agent should have this primitive and that's why we've open sourced the spec. It's lightweight and straightforward to incorporate in any harness!
Introducing Dreaming: Across sessions, Devin builds a memory graph of how you like to work At night Devin self-improves its memory to remove stale records and discover latent information We are creating an OSS standard called Agent Memory Repo: cognition.com/agent-memory-r…
22
19
1
213
586,675
Adhyyan Sekhsaria retweeted
We're also open-sourcing our memory format: - support graph relationships - updates over time with historical records - backed by git and markdown and you can use it with any agent, not just Devin!
Introducing Dreaming: Across sessions, Devin builds a memory graph of how you like to work At night Devin self-improves its memory to remove stale records and discover latent information We are creating an OSS standard called Agent Memory Repo: cognition.com/agent-memory-r…
104
113
16
2,473
197,357
We've spent the past few months iterating on the nuances of this internally. Initially we saw memories degrading Devin runs because they used to store over-generalized learnings. Dreaming was able to fix a lot of this!
Introducing Dreaming: Across sessions, Devin builds a memory graph of how you like to work At night Devin self-improves its memory to remove stale records and discover latent information We are creating an OSS standard called Agent Memory Repo: cognition.com/agent-memory-r…
7
5
2
81
6,416
Adhyyan Sekhsaria retweeted
Introducing Dreaming: Across sessions, Devin builds a memory graph of how you like to work At night Devin self-improves its memory to remove stale records and discover latent information We are creating an OSS standard called Agent Memory Repo: cognition.com/agent-memory-r…
149
216
155
2,587
637,905
Adhyyan Sekhsaria retweeted
agent editors & coding agents ranked. (will post a detailed breakdown)
41
4
6
134
27,184
Adhyyan Sekhsaria retweeted
new type of mind just dropped
Replying to @BenKoska
turns out, it actually kind of works? despite not having been trained on natural language or natural language reasoning, it was able to produce kernels that passed all tests 28% of the time and produced speed ups of up to 66x compared to triton kernels
7
15
238
18,273
Adhyyan Sekhsaria retweeted
you can also use Jev to cut costs and token usage! make your agents more efficient by compacting shell and large file output before it reaches the LLM passing only necessary tokens to the model and carrying fewer tokens through the rest of the conversation
33
15
6
220
24,623
Adhyyan Sekhsaria retweeted
Kev-0.5B: A tiny open source Jev-like decision model with a TypeSafe-compatible API based on Qwen2.5-0.5B that you can train and run on a MacBook Pro. Model card and weights are available on GitHub github.com/jaredpalmer/kev
91
204
72
2,599
416,710
Jev + prompt caching would cut costs by 10x more Too cheap too meter!
Jev’s context window is only 32k tokens while a Claude conversation can reach 200k tokens or more! For compaction, we need to get creative in how to fit up to 200k tokens in the state How it works:
1
26
3,018