Het Patel

Building (AI Software Engineer in the cloud)

[Het Patel @HetPatelx]

Introducing ytimesy, an AI software engineer in the cloud.

ytimesy scores 58.5% on Terminal-Bench 4 with GPT-6 Astra (xhigh) vs Codex 57.9%, same model.

> Use your ChatGPT, Claude or Grok plan
> Continue your Claude Code, Codex or Cursor sessions in the cloud
> Parallel subagents, each on its own machine with its own terminal, desktop and browser
> Memory + Dreaming: remembers your project's rules and decisions across chats, and sharpens that memory while you're away

[Het Patel @HetPatelx]

@Mauricedv Same ones (I used Haiku vs GPT 6.1 sol for comparison).

[Het Patel @HetPatelx]

"Done" is the biggest lie coding agents tell.

If you use one, you've seen it. The agent says done, its own tests pass, and the thing is still broken.

I spent a week trying to fix it at the harness level, measuring every attempt on Terminal-Bench.

First, a checker agent:
→ A second agent that never sees the builder's reasoning
→ Gets the request and the finished work, writes and runs its own tests, sends findings back
→ Has to cite commands it actually ran
→ Works out the answer before reading the builder's
→ Finally, writes its tests blind, before seeing the work
→ Also ran on a model from a different lab (Claude Haiku) to rule out a model blind spot

Result on Terminal-Bench 4.0:
→ Saved runs: caught 6 of 10 failures
→ Live: about 1 in 4, at 75% more cost
→ Haiku: 1 of 5

Then, a requirement ledger, since the checker's misses were misreadings:
→ Quote every requirement word for word
→ List every way each one could be read
→ Build one answer that satisfies all of them
→ Every requirement's check has to run before finishing

Result on Terminal-Bench 2.1:
→ Hard tasks: 15 of 36 vs 8 of 36
→ Fresh tasks: 6 of 30 vs 7 of 28

What I learned: every one of these fixes asked a model to judge the work. When the agent misreads the task, any model reading the same request tends to misread it the same way, so it confirms the mistake instead of catching it.

[Het Patel @HetPatelx]

https://t.co/qrkj5hcnZi

[Het Patel @HetPatelx]

The harness is open source:
https://t.co/krTFoakCpF

[Het Patel @HetPatelx]

@chenzhangreal Haha, the whole point is to demonstrate how strong the harness is, not to claim that ytimesy beat Codex.

[Het Patel @HetPatelx]

Introducing ytimesy, an AI software engineer in the cloud.

ytimesy scores 58.5% on Terminal-Bench 4 with GPT-6 Astra (xhigh) vs Codex 57.9%, same model.

> Use your ChatGPT, Claude or Grok plan
> Continue your Claude Code, Codex or Cursor sessions in the cloud
> Parallel subagents, each on its own machine with its own terminal, desktop and browser
> Memory + Dreaming: remembers your project's rules and decisions across chats, and sharpens that memory while you're away

[Het Patel @HetPatelx]

https://t.co/o8G35ET4aH

[Het Patel @HetPatelx]

https://t.co/qrkj5hcnZi

[Gregor Zunic @gregpr07]

Breaking: Browser Use + Jev = Ultrafast ⚡

Findings flights took 7s and cost only $0.0039 🤯

> new action space every step
> DOM state space
> small LLM fallback to type

(this video is at 1x speed btw)

Built a tiny open source browser agent. try it below ↓

[Het Patel @HetPatelx]

new agent-hop update

use it as a runtime for your coding agents.

new tabs. new workspaces. all inside the terminal.

you’re in Claude, you hop, Codex has the thread.

100% open source ↓
https://t.co/6AMXiBZIyf

[Het Patel @HetPatelx]

https://t.co/6AMXiBZIyf

[Het Patel @HetPatelx]

Introducing agent-hop v0.1.0 — hop between harnesses (not models) mid-chat. entirely rewritten in rust.

Claude Code → Grok → Pi → OpenCode → Codex

the runtime for your coding-agent harnesses

[Gregor Zunic @gregpr07]

Introducing: macOS Harness. A persistent Python harness for virtually any Mac task ⌘

Control your Mac through the accessibility tree, AppleScript, screenshots, or raw coordinate input.

> Full macOS control
> Browser + files + shell
> Virtual cursor - your mouse stays yours
> One persistent python process

It vibe-launched itself on GitHub + PyPI and drafted this tweet without me in the loop 👀

100% open source 🔗↓

[Het Patel @HetPatelx]

Try it now ↓
https://t.co/5JjIKq41nV

[Het Patel @HetPatelx]

Continue your Claude Code chat in Codex — with attachments and tool calls too!

agent-hop:
> Search across every coding-agent chat — no more cd-ing around to find and resume the right one
> Resume in Claude Code, Codex, OpenCode, Pi, or Grok Build — any agent, not just the one you started in
> Real conversation, tool calls, and attachments carried over — not just text

[Het Patel @HetPatelx]

@catmanyau https://t.co/42PEAvkwZC

[Het Patel @HetPatelx]

I kept losing useful AI coding chats across Claude Code, Codex, OpenCode, Pi, and Grok.
So I built agent-hop.

It searches all your local coding-agent sessions, then lets you continue from any chat context in any agent.

A Codex chat can become an OpenCode chat.
A Claude chat can become a Codex chat.

> npm install -g agent-hop
> agent-hop

[Het Patel @HetPatelx]

@FlnBGNfiber everything is local

[Het Patel @HetPatelx]

hit a limit in one coding agent, keep the exact same conversation going in another.

agent-hop resumes any of your coding-agent sessions (Claude Code, Codex, OpenCode, Pi, Grok) in a different agent, with the real conversation history, not a summary. it also searches across all of them, so you never have to remember which agent or project a chat happened in.

here's a demo of Claude Code → Codex, same conversation, full history, zero re-explaining.

npm i -g agent-hop

[Het Patel @HetPatelx]

https://t.co/6AMXiBZIyf

[Het Patel @HetPatelx]

https://t.co/6AMXiBZIyf