
A cool guide to run multi agents with a shared memory for development + mcp resource
[supprimé]

[supprimé]
I work a lot with browsers when it comes to giving visual context to LLMs.
The usual workflow was:
take a screenshot → upload it to my IDE → prompt the context.
That works fine , until you’re clicking 1,000 screenshots a day. Eventually they pile up in storage, and ironically, storage costs keep skyrocketing.
So I decided to make my life easier.
I built agent-vision
Github Repo -> https://github.com/kedarvartak/agent-vision
NpmJS Package -> https://www.npmjs.com/package/agent-vision-mcp
agent-vision is a vision layer between your development environment and your browser. It gives LLMs live browser context - not just screenshots or layouts, but:
URL
DOM
Element attributes
Network events
Viewport size
Console logs
Tab title
No more constantly switching tabs for browser tasks.
I work a lot with browsers when it comes to giving visual context to LLMs.
The usual workflow was:
take a screenshot → upload it to my IDE → prompt the context.
That works fine , until you’re clicking 1,000 screenshots a day. Eventually they pile up in storage, and ironically, storage costs keep skyrocketing.
So I decided to make my life easier.
I built agent-vision
Github Repo -> https://github.com/kedarvartak/agent-vision
NpmJS Package -> https://www.npmjs.com/package/agent-vision-mcp
agent-vision is a vision layer between your development environment and your browser. It gives LLMs live browser context - not just screenshots or layouts, but:
URL
DOM
Element attributes
Network events
Viewport size
Console logs
Tab title
No more constantly switching tabs for browser tasks.
As a software engineer, you eventually become QA for your own features too. At first it’s manageable. Then you realize you’re running the same regression flows every deployment.
So I built Easy Sanity — an autonomous AI agent for regression & sanity testing.
You define test cases in plain English, and the agent executes them for you.
It supports two modes:
• Deterministic flows → predictable, fast, low-token execution for stable regression suites
• Adaptive flows → browser-native reasoning for dynamic UIs and exploratory validation
For adaptive workflows, Easy Sanity uses Browser Harness by browser-use to connect agents directly to live browser sessions through CDP instead of repeatedly generating Playwright scripts.
The goal is simple:
Focus on shipping. Let the agent handle the sanity.
Repo: https://github.com/kedarvartak/easy-sanity
Docs: https://kedarvartak/agents/easy-sanity
Linkedin: https://www.linkedin.com/posts/kedar-vartak20\_as-a-software-engineer-you-eventually-end-activity-7458905107645009920-3aVL?utm\_source=share&utm\_medium=member\_ios&rcm=ACoAAEIgu4gB04kNAIaHNXANbvZjcB\_mJNsj\_L4
As a software engineer, you eventually become QA for your own features too. At first it’s manageable. Then you realize you’re running the same regression flows every deployment.
So I built Easy Sanity — an autonomous AI agent for regression & sanity testing.
You define test cases in plain English, and the agent executes them for you.
It supports two modes:
• Deterministic flows → predictable, fast, low-token execution for stable regression suites
• Adaptive flows → browser-native reasoning for dynamic UIs and exploratory validation
For adaptive workflows, Easy Sanity uses Browser Harness by browser-use to connect agents directly to live browser sessions through CDP instead of repeatedly generating Playwright scripts.
The goal is simple:
Focus on shipping. Let the agent handle the sanity.
Repo: https://github.com/kedarvartak/easy-sanity
Docs: https://kedarvartak/agents/easy-sanity
Linkedin: https://www.linkedin.com/posts/kedar-vartak20\_as-a-software-engineer-you-eventually-end-activity-7458905107645009920-3aVL?utm\_source=share&utm\_medium=member\_ios&rcm=ACoAAEIgu4gB04kNAIaHNXANbvZjcB\_mJNsj\_L4
Since TPO isn’t transparent and statistics are shit. Plan ahead - company trends follow the same pattern
Note - Yet to get 24-25 batch data