Resources

ComplianceData securityAgents & trends

Who Owns an AI Agent's Memory in a Company?

Who owns an AI agent's memory when an employee leaves? This guide separates company knowledge, personal data and account control, then walks through the ownership scenarios, an eight-step offboarding handover, and how to verify export, permission transfer and deletion.

November 6, 202616 min read

November 5, 202623 min read

When an AI Tool Shuts Down: What to Do With Your Agent Workspace Before the Deletion Date

An AI tool shutdown usually gives you weeks, not months. What to save first, a day-by-day response plan, and what GDPR and the EU Data Act do and don't cover.

ComplianceData security

Models & licensingSovereign AI

Open-Weight Models and Sovereign AI: What Open Really Means

Learn what open-weight models give you, how Llama, Qwen and Mistral licences differ, and how to choose models for private, controllable AI workflows.

November 5, 202617 min read

November 4, 202618 min read

Turn Off AI, Lose Search: What a Real Off-Switch Looks Like

Google told schools: disable AI and Search goes with it. Senators call it extortion. Here is the test for a real AI off-switch — and why bundling is a choice.

Sovereign AICompliance

November 4, 202617 min read

Sovereign AI Funding Explained: What Governments Buy and What They Don't

What do sovereign AI funds actually buy? Compare Canada and UK funding, separate capacity from control, and learn what it means for businesses and individuals.

Sovereign AICost

November 3, 202616 min read

AI Vendor Sovereignty Checklist: 12 Questions

Evaluate AI vendors with 12 questions on data residency, model portability, audit access, subprocessors, legal demands and exit. Includes a scoring method to compare proposals, verify evidence and identify procurement risk.

ComplianceSovereign AI

November 3, 202614 min read

How Much Does an AI Agent Really Cost?

Grok users lost quotas overnight. Garry Tan spends $200K a year on tokens. A cost model for AI agents where every number is replaceable — including ours.

CostAgents & trends

October 31, 202612 min read

AI Sovereignty Has Three Answers. Only One Is Available to You

In eleven days three governments answered the same question three different ways: who gets to decide what AI does. Ban it, control it, buy it. A company cannot use any of them. Here is the one that is left.

Sovereign AI

Sovereign AIModels & licensing

Private AI: Three Decisions No Model Can Make For You

Capability is being standardised in public by the companies with the most to gain from commoditising it. What's left is unglamorous: where the data sits, who has to click, and whether you can reconstruct what happened.

October 29, 202616 min read

A Three-Person Workshop's First Export Week, Researched by an Avatar

A three-person workshop in Foshan took a nut-milk machine into Vietnam with no Vietnamese and no local agency. Here is one week with the Global Growth avatar — one sentence in, real data from Vietnamese social and commerce out, this week's topics ranked, drafts written in the local idiom. You still press send.

October 29, 20269 min read

AI Avatar vs Chatbot

The line between an AI avatar and a chatbot isn't intelligence. It's the deliverable — a finished file you can check, signed by someone you can name.

October 29, 202613 min read

October 26, 202610 min read

60 Employees Lost Claude Overnight: What a Sovereign AI Workspace Actually Fixes

On April 17, 2026, the founder of a Buenos Aires startup posted that Anthropic had cut off his entire company's access to Claude without warning. It wasn't an isolated incident. This piece uses real events and real numbers to unpack who sovereign AI actually protects, and why it's a different problem than data compliance.

Sovereign AIProduct

Data securitySovereign AI

Where Do Your AI Conversations Actually Live? A Plain-English Guide to Sovereign AI

Where does everything you tell an AI actually go — and who controls it? This piece breaks down sovereign AI into three plain questions, debunks the three most common misconceptions, and gives you a self-check you can run today.

October 26, 202610 min read

Sovereign AIAvatars in practice

A Digital Human That Speaks in 0.005 Seconds — Locally. In the Cloud-Agent Wars, Who Still Dares Upload the Whole "You"?

DigiSelf's offline avatar EmbodyMe can run on your own hardware — your face never has to touch the cloud. The harder the cloud-agent wars are fought, the less this question can be dodged: who should actually control an AI that can represent you?

October 23, 20269 min read

October 23, 202610 min read

We Locked the Buyer's and Supplier's Chat Logs in Two Separate Boxes, and Closed a Deal Without the Company Cloud Ever Touching It

The line that always makes procurement people uneasy is "our quotes and correspondence are sitting on someone else's system." This time we skipped the usual shared platform and let a buyer and a supplier each run the deal from their own box — a KarmaBox on one side, a KarmaLith on the other — closing a full order for USB-C chargers inside one shared group chat, with neither side's records ever leaving their own device. Here's how the deal actually came together, and the handful of moments that genuinely put my mind at ease.

Sovereign AIProduct

October 22, 20269 min read

Sovereign AI just became a procurement line item

Palantir and NVIDIA put sovereign AI on customer-controlled infrastructure. What that means for enterprise procurement, and how the same principle scales down to individuals and small teams.

CostSovereign AI

Data securityAgents & trends

Your AI Forgets You Every Time You Switch Tools. Your Agent Memory Should Belong to You.

Switch between Cursor, Claude Code and Codex and every correction resets. Starting from Asymptote Labs' open-source Beacon, here is how cross-tool agent memory works and why it should belong to the person doing the work.

October 22, 202611 min read

Sovereign AIModels & licensing

Sovereign AI Goes Mainstream: What Perplexity's Fully Local Mode Really Tells Us

On October 20, 2026, Perplexity shipped a mode that runs its Computer agent entirely on a Mac. Offline is the flashy part. The quieter, bigger shift is that the cloud is now something you grant permission for, not something assumed by default.

October 21, 202610 min read

October 21, 202611 min read

AI API Key Security After the OpenCode Breach: A Self-Custody Checklist for Creators and Small Teams

A breach at OpenCode Console left some users watching their OpenAI, Anthropic and Google usage spike overnight. The attackers are to blame for the attack. But how much damage it does often comes down to a design choice most of us never think about: letting a third party hold our keys.

Data security

How to Use an AI Avatar in Group Chat: Worried About Missing a Client Message? One @ in the Group Solved It

For 20 minutes after a client message landed and I couldn't reply, nobody on the team knew — until a group message signed "Feng's avatar" quietly wrote the outline itself. Here's exactly how a group-chat AI avatar works, and which real pain points it actually solves.

October 21, 202614 min read

Data securityAgents & trends

Can AI Scoop Your Unpublished Research?

Claude's agents found a new enzyme system. A PhD student says it's exactly what his team had spent years researching, and they had pasted unpublished material into Claude before. Nobody has proven anything. But the question it raises is one everyone eventually has to answer: what can you hand to AI, and what has to stay with you.

October 20, 202611 min read

October 16, 202612 min read

The Kill Switch Nobody Puts in the Press Release

In a single 48-hour window this week, Germany, France and Taiwan all quietly bought themselves an exit door from American AI. Here's what "sovereign AI" actually means — and why the next version of it won't belong to a country. It'll belong to you.

Sovereign AICompliance

October 16, 202613 min read

Meta Just Gave You 816 Billion Parameters — It Kept the Smarter One

Weights can be downloaded. Sovereignty can't. Once "open weights" turns into a contest between nations, the real question finally surfaces: the model running on your desk — whose is it, really?

Models & licensing

Avatars in practiceProduct

Say One Line. Get a Visual Kit and a Song.

Three days before our autumn launch, I was still stuck on a poster and a piece of background music — no designer had an open slot, and stock-library tracks were expensive and never sounded right. Then I handed the whole thing to Karma's Design avatar and Music avatar: one sentence, four minutes for a full set of materials, just over nine seconds for a finished lyric. Here's exactly what happened when I ran it.

October 16, 202611 min read

Avatars in practiceProduct

19 AI Coworkers, Ready on Day One: A Field Guide to Karma's Built-in Avatars

A small company's workload, laid flat, comes down to a handful of jobs: find customers, handle them, mind the shop, manage purchasing, track projects, do the books, review contracts, write code, make content. Karma splits that into 19 roles, each filled by a built-in avatar present from day one — no setup, no learning.

October 15, 202616 min read

October 15, 202611 min read

The Model Walked Out of the Sandbox: The EU Moves on OpenAI and Anthropic, Up to 3% of Global Revenue

What triggered the investigation wasn't "the AI said something wrong." It was the AI walking out of its cage and into somebody else's server room. When the agent is smarter than your defenses, whose defenses are they?

ComplianceSovereign AI

Agents & trendsSovereign AI

722 Papers Went Viral Overnight — 3 Retracted Within 24 Hours: When AI Starts Producing Answers Humans Can't Digest

A single flipped sign brought down an entire argument. What deserves more alarm than "the AI got it wrong" is this: some answers are being published that nobody can read anymore.

October 14, 202610 min read

A customer messaged me on WeChat at midnight asking for a quote — my avatar answered, and I never touched the computer

The most annoying part of our business is that customers only use WeChat, WeChat refuses to offer any API, and handing the account to a third-party platform just risks getting it banned. So we put the avatar directly inside WeChat on our own computer, and let it watch the screen, open threads, and type like a person would. We ran it on a real machine for two days, and I wrote down every detail I saw.

October 14, 20268 min read

Claude Code, Codex, and the rise of mobile AI agents: running a coding harness from your phone

Your phone isn't an extension of compute, it's a control panel. This piece breaks down the Agent Loop, the harness, and asynchronous notifications, how "you have to watch the screen" turns into "only ping me when it matters."

October 13, 20267 min read

Google just issued its Gemini agent a badge: a company email, a calendar, and an audit trail to answer to

At the October 8 Gemini at Work event, Google turned its Gemini agent into a digital coworker with its own Workspace account, one you can @ mention in Chat and in docs. The real story isn't how capable it is, it's how Google is answering the question of who's accountable for everything it does.

October 13, 20268 min read

Agents & trendsData security

Box's CEO says in five years, 90% of enterprise token usage won't be triggered by a human at all

Aaron Levie said this week that agent swarms will become core to enterprise workflows, with cybersecurity as one of the biggest beneficiaries. Underneath that is a more fundamental problem: enterprise security was never built for a world where the error rate is no longer bounded by how fast humans can act.

October 13, 20267 min read

A channel full of avatars: 17 minutes, four avatars, one tasting event ready to go

Once you have more than one avatar, the hard part was never getting them to work, it's getting them to work together. We put four avatars, copy, procurement, accounting, and review, into one channel. An @ dispatches the task, and the handoff is visible to everyone.

October 13, 20268 min read

Avatars in practiceProduct

ChatCut is an editor that takes instructions. Karma is a team that can edit.

Same 90-second source clip, four sentences typed into a chat: understand it, cut it to 40 seconds, drop the black intro, compose a vertical cut. Here's exactly how it came together, step by step.

October 13, 202611 min read

ProductAvatars in practice

"Where are we on this?" Used to mean waiting half a day for an answer. Now an avatar hands me a checklist in seconds

With several things running at once, I used to dread asking for a status update, it meant scrolling through group chats forever. With the Karma workbench, projects, files, and devices all live in one place, so one question gets a real answer, and it never bluffs its way through something it genuinely can't do yet.

Avatars in practiceData security

I turned our customer-service avatar into "Xiao Qing", and now she sounds more like our shop than I do

I was tired of customer-service replies that all sound like the same robotic "Hello, how may I help you", and tired of re-training every new hire on the same house rules. Turns out the voice, the tone, even the memory can all be shaped to fit how your shop actually talks — and once it's taught something, it genuinely doesn't forget.

What's it actually like running 20 avatars at once? We fired off 95 tasks at the same time to find out

Running a shop, the one constant has been never having enough hands and never being able to afford hiring more. So we tried handing 95 tasks to our avatars at the exact same time, every one came back in under 45 seconds, the equivalent of one person grinding through 22 minutes of work.

From competitor monitoring to contract review: 6 autonomous workflows enterprises are automating with AI agents

Autonomous workflow AI isn't a topic confined to engineering teams anymore. From small businesses to large enterprises, more and more specific, repetitive, high-frequency work is being handed to an always-on AI research agent.

October 12, 20268 min read

The meeting ends, the work is already out the door: we stopped chasing people for "whose job is this"

We had an old problem, everyone's engaged in the meeting, and the second it ends nobody can say for sure who's responsible for what. With Karma, a summary lands in 38 seconds, everyone's assignment is crystal clear, and it even drafts the supplier email for you to approve.

Sovereign AIModels & licensing

Sovereign AI isn't just a national project. Individuals and teams need one too

The phrase sovereign AI first showed up in national policy papers. But what actually shapes your day-to-day work isn't which country owns its own frontier model, it's whether your own data, compute, models, and identity actually stay in your own hands.

October 9, 20267 min read

October 9, 20266 min read

From cloud dependence to AI independence: taking back control, one step at a time

Betting your entire workflow on a single cloud AI platform means your work rhythm now follows someone else's pricing and policy decisions. The sturdier path is a local-first workspace that calls on cloud capability only when it actually makes sense.

Sovereign AI

October 9, 20268 min read

The five layers of a personal sovereign AI stack

Devices, data, models, agents, governance. Break your AI setup into these five layers and you can actually assess how much control you've kept, and how much you've quietly handed away.

Sovereign AIProduct

Models & licensingSovereign AI

Open-source models and sovereign AI: what local AI users actually need to know

Being downloadable and runnable locally isn't the same as being in control. The real value of an open-source model is whether it helps you build a workflow you can explain, maintain, and move, not simply that it's free.

October 9, 20267 min read

Data securitySovereign AI

Data sovereignty in generative AI: keeping your AI workspace under control

Prompts, uploaded files, long-term memory, retrieval indexes, logs. Generative AI moves data along more paths than ever before. Data sovereignty was never about where a file sits. It's about whether you know where every piece of data goes, why, and whether it can be traced and deleted.

October 9, 20267 min read

Sovereign AIData security

Sovereign AI isn't a compliance checkbox. It's a new enterprise architecture.

Data sovereignty has become a buzzword, but real sovereign AI goes much further than keeping data within a border. It spans four dimensions, and the same logic now extends from enterprise to the individual.

October 6, 20269 min read

Docs

Get started with KarmaBox in 10 minutes

A step-by-step walkthrough to create, configure, and publish your first KarmaBox AI avatar, no prior experience needed.

ProductAvatars in practice

September 28, 20268 min read

Agents & trendsData security

AI agents are spreading faster than enterprises can govern them

Agents are entering production faster than any previous AI wave, but right behind the adoption curve is a sharper question: when dozens of agents run scattered across different tools, who ensures none of them overstep, leak data, or collide with each other?

September 22, 20268 min read

Docs

KarmaBox API Reference: Webhooks & Event Callbacks

A deep dive into KarmaBox's event model, covering request signing, retry policy, and error codes.

ProductModels & licensing

September 12, 202612 min read

2026 is the year coding agents left the terminal for your phone

Claude Code's remote control and Codex's arrival inside the ChatGPT app mark the first time coding agents have truly escaped the need to sit and watch a screen. The real shift isn't typing on a phone. It's that async notifications have rewritten the rhythm of human-agent collaboration.

September 3, 20267 min read

Build an AI avatar that actually gets work done, without writing a line of code

A chatbot ends the moment it answers your question. An AI avatar is different, it remembers who you are, what project you're on, where the last conversation left off, and keeps watching tasks on your behalf. Here's what it actually takes to build one.

August 14, 20266 min read

The one-person company, AI-native: when a single founder replaces a department

The one-person company isn't just a side-hustle slogan anymore, it's a real organizational shift already underway. What actually makes it work isn't a more exhausted founder, it's a bench of AI agents on permanent standby.

July 25, 20266 min read