Kboom

Memes, Cricket, Arsenal Fan, New technologies & Corp law, Capital and Crypto Markets.

Collator of interesting threads under #K7Threadalert

[Sanjeev Kumar @mishrak_sanjeev]

Today I'm open-sourcing AgenticOrg.

Apache 2.0. The whole platform. No enterprise paywall, no bait-and-switch tier.

AgenticOrg is not a copilot or a chatbot with tools. It's an AI workforce — agents that make real, authenticated API calls into your production systems (GSTN, Tally, Zoho Books, SAP, Salesforce, HubSpot, Jira, Slack, Stripe…) and a human approves every critical decision before it executes.

Live in production right now (numbers fetched from /api/v1/product-facts, not hand-typed):

→ 26 production-grade agents across Finance, HR, Marketing, Operations, Back Office, Communications
→ 53 native connectors with 320 typed tools
→ 1000+ extra integrations via Composio
→ Industry pack for Indian CA Firms — multi-client GST/TDS/MCA filing, partner dashboard, encrypted credential vault
→ India-first: GSTN, EPFO, Income Tax, Tally Prime, Darwinbox, Account Aggregator, Adaequare GSP

Why this matters:

1. Every new agent starts in SHADOW MODE. It processes real data but takes no action. It must pass 6 quality gates — accuracy, safety, performance, reliability, security, cost — before promotion. You see exactly what it would have done, before you let it.

2. Every API call goes through the Tool Gateway. Idempotency keys (Redis-backed, 24h TTL), token-bucket rate limits, automatic PII masking via Microsoft Presidio. No duplicate transactions. No runaway loops. No leaked Aadhaar / PAN / GSTIN.

3. When confidence drops below 88%, the agent stops and asks a human. Every approval has a deadline. Every decision is HMAC-signed in the audit trail with approver, timestamp, and notes. This isn't a "human-in-the-loop toggle" — it's a governance engine.

4. Multi-tenant by construction: Postgres Row-Level Security, Grantex offline JWT verification with manifest-based scope enforcement on every tool call, server-side tenant binding (never trust client tenant_id).

5. Smart LLM routing via RouteLLM across 3 tiers (Gemini Flash → Gemini Pro → Claude Sonnet, configurable to Opus / GPT-4o), per-agent budget enforcement with auto-block on overspend, and an air-gapped path via Ollama / vLLM for regulated workloads.

That's the difference. Not "AI that can do things." AI you can actually trust to do things — and prove it after the fact.

Built in the open:
🐍 pip install agenticorg
📦 npm i agenticorg-sdk
🔌 MCP server in the public registry — works with Claude, ChatGPT, Cursor, any MCP client
🤝 A2A protocol with .well-known/agent.json discovery — your agents are callable by any A2A-compliant system
🧠 LangGraph orchestration with conditional branching, parallel teams, dynamic re-planning when steps fail
📝 Upload an SOP (PDF/DOCX/Markdown) → get an agent. CSV org chart → get an agent hierarchy. No prompts to write.

GitHub: https://t.co/5HK5cnG4X8
Live demo: https://t.co/87cuvChgJ3
Playground: https://t.co/CmetrCXDjx

Stars and forks make it real. 🌟

#OpenSource #AgenticAI #AIAgents #LangGraph #MCP #Enterprise #BuildInPublic #IndiaAI

[Eric Glyman @eglyman]

Introducing Stack.

The AI operating system that lets accounting firms take on more clients without hiring. Learns your firm's process, runs the close, posts the journals. Fully auditable.

We’re living through the biggest shift in accounting since the spreadsheet.

[Kboom @vrk7here]

@charliejhills stack

[Higgsfield AI 🧩 @higgsfield]

Higgsfield plugin for Figma is live.

Generate images with every model, create vectors as clean SVGs, build mockups in Mockup Studio.

Cut out backgrounds, apply color grades. Swap faces while keeping the scene, shoot studio product photos with no camera, animate hero creatives into ads.

[LM Studio @lmstudio]

Meet LM Studio's mobile app.

Your local models, now in your pocket.

[Arjun Mahadevan (Mr. LLC 🇺🇸) @ArjunMahadevan]

The gap between "I have an idea" and "I have a company" used to be weeks of paperwork.

With @doolaHQ inside @perplexity_ai Computer, it's one prompt.

Proud to be a launch partner for the Growing Businesses bundle today.

Try it at: https://t.co/yHorS1IfUR

[Nous Research @NousResearch]

The next evolution of Hermes Agent is here!

Introducing Hermes Desktop: everything you love about Hermes, now native on your machine.

First demoed in Jensen's GTC keynote, it's now in public preview.

[Sachin Tendulkar @sachin_rt]

Congratulations to @RCBTweets on winning consecutive @IPL titles. One of the challenges in sport is that success changes the questions you have to answer. After winning once, the task is no longer proving you can do it, but proving you can sustain it. RCB met that challenge impressively this season. A well-deserved achievement!

[Sushant @sushantpandey_]

Introducing Prava- a stack for agentic checkout on AI Apps

Supports Card & Wallet

Check docs, get early access!

[The Khel India @TheKhelIndia]

🚨 DEV MEENA BREAKS THE NATIONAL RECORD!

- Great 5.42m jump in Men's Pole Vault 🔥🔥🔥

WOOOOOAH, WELL DONE DEV! 🇮🇳💙

[H S Phoolka @hsphoolka]

2nd Flying Sikh is BORN! Gurindervir Singh has shattered the national record, 100M in 10.09 seconds.Very close to Usain Bolt’s 9.58s world record, Just half second away from history! Congratulations champion. Waheguru ji kripa rakhan sher putt te! #GurindervirSingh #FlyingSikh

[Kboom @vrk7here]

@Dork_sense @morphic Any walk through videos in your YT that you could share, to use this feature ?

[CyrilXBT @cyrilXBT]

https://t.co/U0R2au5Zka

[Dorksense @Dork_sense]

i've been building @morphic workflow. And this is absolutely crazy.

https://t.co/G7gc49He92

What are morphic workflows: i build a full creative pipeline and save it as a workflow. each step can have its own model yet chained with its own prompting strategy, and dynamic inputs that shape how the pipeline runs. i can define how i think about a task, bake my prompts and assets into it, and the workflow carries that forward every time.

The brain: each step can carry its own prompting strategy. how you prompt an image model is different from how you prompt a video model, and the workflow can hold that knowledge per step. and i can use a workflow as context to create other workflow plus if i find a better approach, i update that step. and that just works

Routing: try doing conditional routing on a node graph without it turning into spaghetti. on morphic, i set up inputs that change the entire pipeline path. i have 1 workflow which continues to run until nano banana pro doesn't stop flagging my generation ;)

Assets: i can save media directly inside workflows. reference images, character sheets, audio, video clips, scripts. if i'm working on a film, character A and character B reference photos live inside the workflow. the pipeline knows which character is which. if i always work at 2K res with nano banana pro, that's saved too. i don't need to re-attach or re-explain anything every run.

Non-AI tools: morphic has QOL utility tools like frame extraction, trim, resize, speed change, compose. these can be steps inside a workflow alongside the AI ones. need to extract a frame mid-pipeline before the next generation step? the utility tool handles it inside the same chain.

Model swapping: nothing is locked. new model drops, swap it into any step. same workflow, different engine.

and there is still more to explore 😭 Truly Wires without wires

[Kboom @vrk7here]

@morphic awesome 😀😀

[Ravi Kiran S @vikataravi]

A breakthrough for Indic manuscript OCR.

Introducing UniLipi - a unified Sanskrit manuscript OCR system capable of decoding across 13 scripts.

[AI Edge @aiedge_]

Google is using AI to cure cancer.

This is literally the craziest AI use case I've ever seen.

What I just discovered in the London Google DeepMind offices will blow your mind:

[Kboom @vrk7here]

@charliejhills saving this, thank you!

[Kevin Simback 🍷 @KSimback]

https://t.co/dCgExZ0et9

[Irfan Ali @iiamirfanali]

चलती Train के दरवाजा पर बाबा थे और सबकी सांसें अटकी हुई थी। Door अंदर से लॉक था, फिर Police ने आकर दरवाज़ा खोला। आखिर बाबा की मंशा क्या थी? 😎