Ashish

AI Coding Harnesses & Agentic Workflows

[Ashish @BuiltinMind]

Ten Grok Bot templates for the actual founder loop: survive, decide, ship, review, sell, post. Plus a shared memory repo so the whole org gets smarter at once, not ten private diaries.

A pile of bots is not a startup. A pile has no graph, no memory. https://t.co/igEVrUlOKv

[Ashish @BuiltinMind]

Finally the harness from most cracked team on the planet

[DeepSeek Harness @DeepSeekHarness]

DeepSeek Harness for desktop is now available on macOS and Windows. https://t.co/hpDsyMYvLs

[Ashish @BuiltinMind]

@klashsarma https://t.co/oW6ArYf5Tb

Lets see how it works

[Ashish @BuiltinMind]

https://t.co/RpFOErJqE5

[Ashish @BuiltinMind]

Paper: SemNav: Semantic Navigation for Repository-Level Issue Localization

Your coding agent is hunting the files that matter for a bug report. The repo is huge. It keeps opening whole source files and guessing.

SemNav is a harness for that hunt (the code around the model that decides what it can open, how it jumps between symbols, and which suspects stay on the list).

Ordinary search seeds a broad candidate list. The agent revises that list as it learns more.

Go-to-definition jumps to the real callee across files instead of searching among name collisions.

Short issue-conditioned cards summarize what a function does for this bug, with a few supporting snippets, so the agent does not have to keep the whole file in working memory.

A candidate workspace keeps each suspect with the evidence that still justifies it.

Same Gemma 4B model on Software Engineering Benchmark Lite (SWE-bench Lite).

Share of issues where the correct file appears in the top 10 guesses:

Strongest prior: 68.33%.
SemNav: 82.67%.

Those better file picks feed the repair step. With the same localization model feeding a fixed repair backend, resolved issues move from 44.00% to 52.33%. Most of the lift is the localization harness, not a bigger writer.

[Ashish @BuiltinMind]

@ataiiam Got it makes sense

[Ashish @BuiltinMind]

@sunjiao123sun_ @LyalinDotCom When can we expect it in antigravity ?

[Ashish @BuiltinMind]

@TheAhmadOsman @victormustar Seems like all chinese labs are just postponing releases since Astra and Opus 5.5.

Sol 6.1 and Gemini 4 Argon are the only model relevant on pareto frontier with opensource

[Ashish @BuiltinMind]

@orionweller @LyalinDotCom @GoogleDeepMind Haha Non Flash model

Irony is everyone is now waiting for gemini 4 flash

[Ashish @BuiltinMind]

This is significant breakthrough in context engineering

Basically turns context into active learnings stored and fetched from memory on priority and relevance.

Need a detailed walkthrough on this

[Rulin Shao @RulinShao]

‼️The Bitter Lesson for context management: Giving LMs unrestricted control over their context beats human-designed SOTA!

Introducing 🩵Context Language Models (CLMs)🩵
- Natively manage their own context
- Treat context as a file
- Learn policies in CLM weights, no harness

[Ashish @BuiltinMind]

Waiting for Gemini 4 on pareto frontier

[Ashish @BuiltinMind]

@AravSrinivas Wow advance Dussehra gift

[Aravind Srinivas @AravSrinivas]

We’re open sourcing our state-of-the-art contextual embedding models, which perform best in turbopuffer’s context-bench.

[Ashish @BuiltinMind]

@Hiteshdotcom unka toh din hai 😉

[Ashish @BuiltinMind]

I was just thinking for pro subscription.

[shimrit ben-yair @shimritby]

We're working hard to bring Gemini 4 Argon to @Google AI Ultra subscribers as soon as possible!
https://t.co/igByQWmRVK

[Ashish @BuiltinMind]

@danshipper Hope they dont nerf it

If Mythos becomes Fable after DoC review Argon becomes Borat ?

[Ashish @BuiltinMind]

@alexandr_wang Loved it hope you keep improving it and release Muse 1.4 its been weeks lol

[Ashish @BuiltinMind]

Muse App Review:
First and formost no company provides Data controls as accessible as Muse. Mark has learned his lessons about data privacy.

[Ashish @BuiltinMind]

@jilyannori Glad you like it

[Ashish @BuiltinMind]

Muse App Review:
First and formost no company provides Data controls as accessible as Muse. Mark has learned his lessons about data privacy.

[Ashish @BuiltinMind]

1M token output limit. Thats insane.
Gemini 4 could write a book in a single response.

[Ashish @BuiltinMind]

@sundarpichai A Monster release