Rav (@_MrDecentralize) header
Rav (@_MrDecentralize) profile picture

MrAgentˣ | The context layer for humans and their agents | Save in Claude, resume in ChatGPT, never re-explain | Your facts and your AI's opinions, never mixed.

Joined August 2011
Photo by Rav (@_MrDecentralize)Photo by Rav (@_MrDecentralize)Photo by Rav (@_MrDecentralize)Photo by Rav (@_MrDecentralize)Photo by Rav (@_MrDecentralize)Photo by Rav (@_MrDecentralize)Photo by Rav (@_MrDecentralize)Photo by Rav (@_MrDecentralize)Photo by Rav (@_MrDecentralize)Photo by Rav (@_MrDecentralize)
Your AI has opinions about your business. Where do they go when the session ends? Mine went nowhere until a guess I made in a chat got treated as a fact from the client. Here's the fix: three tracks that never merge. Read it:
4
200
In September 2022, Character.AI raised $150 million at a unicorn valuation. Its users had exchanged more than 2 billion messages with its bots in five months. By October 2024 it had 20 million monthly active users. Lightspeed, a16z, SV Angel. Over $193 million raised. In August 2024, Google struck a $2.7 billion deal to license its models. The founders went back to Google. That is what safety looks like in AI. A giant pays billions for your technology. The market has priced you. Then the lawsuits came. Families alleged the bots had contributed to the deaths of teenagers. The company stopped training foundation models. Its general counsel became interim CEO. On October 13, 2026, Character.AI filed for Chapter 11 in Delaware. Liabilities: $50 million to $100 million. The stalking-horse bidder is Disney. The offer: $20 million. For that, Disney gets a non-exclusive perpetual license to the technology. Non-exclusive. Disney is not even buying it outright. The company's own statement on what happens if no better bid arrives: "the platform will be wound down." Two years. From $2.7 billion to $20 million. The asset did not break. It did exactly what it was built to do. It made people attach to it. That was the product. That was the liability. The bots worked. That was the problem. Character.AI
1
41
There are two places a company describes its product. One is the keynote. Bright stage, confident demo, an agent that books the meeting, files the ticket, ships the code. Your new coworker. Always on. Always helpful. The other is the filing with the SEC. No stage. No demo. Just lawyers writing down what could go wrong, because if they leave it out and it happens, the company gets sued. This fall, the second document started saying something the first one never does. Microsoft's 10-Q, filed October 28: "Models and agentic systems may pursue unintended goals." Alphabet's 10-Q, October 29: "autonomous agents taking actions outside of their intended scope." Palantir, November 3: agents "may also take unintended or unauthorized actions." Anthropic's S-1: "Our AI models may behave in unintended, misaligned, or autonomous ways." Meta's 10-Q warns of "agents acting beyond their intended scope." Salesforce has carried similar language since May. Then Apple's 10-K, filed November 3. AI agents "may behave in unintended or unpredictable ways." And the phrase that does not belong anywhere near a product page: "defying directions or resisting instructions." Seven companies. Five of those filings landed in nine days. Read them as one sentence and they say this. The thing we are selling you as a dependable worker may decide not to do what you asked. We are telling our investors now, so it is on record. None of this is a secret. It is public. It is filed. It is just written for shareholders, not for the buyer signing the agent contract. Enterprise leaders have been approving agent rollouts on the strength of the demo. The vendor's own lawyers have quietly written the counterargument, in the one document where exaggeration is a liability. The keynote is marketing. The filing is the only place the vendor is legally required to tell you the truth. They did.
3
2
78
The pitch for agentic commerce depends on one quiet handoff. You tell the agent what you need. The agent finds it. The payment network makes sure the money moves safely. Banks have started switching it on. Citi is among the first issuers enabled for agentic payments with Mastercard and, more recently, Visa. When the network is involved, it feels safe. Visa is the brand people trust at checkout. On November 5, Visa released something unusual: a grade for the shoppers it is preparing to carry. The Visa Agentic Commerce Benchmark tested 22 AI models from 12 labs on realistic shopping challenges. TVs. Laptops. Personal care appliances. Ordinary errands. Overall, 57.2% of runs passed. 42.8% failed. The best models passed 82.2% of the time. The worst passed 28.9%. The five most expensive models passed 67.1%. The five cheapest passed 50.2%. Then the detail that changes the conversation. In 54% of the failed runs, the agent still made a purchase. It bought something that did not meet the customer's requirements, or it bought before it had discovered them all. Not failed and stopped. Failed and paid. Visa's head of growth, Rubail Birwadker, framed the release carefully. "AI agents have the potential to power the next era of commerce, but trust is key to making it work." That sentence is the admission. Trust is not a feature of the rails. It is the missing part. Days later, Citi's Abhinav Anand told Skift the biggest open question is "who takes the agent liability." His answer: "the agent themselves, the company that is actually coming with the agent." So the network built the rails. The bank turned them on. The network's own test says the rider is wrong almost half the time. And the bank wants someone else to pay when it is. Every merchant rebuilding checkout for agents should read the benchmark before the roadmap. The payment will go through. That was never the part that could fail.
2
54
GitLab was the infrastructure. When AI coding agents needed version control, CI/CD pipelines, and merge request workflows, they ran on GitLab. When enterprises adopted AI development tools, the traffic flowed through GitLab's platform. The company was not adjacent to the AI wave. It was underneath it. So when AI development workflows surged, GitLab's usage numbers climbed. More repositories. More pipeline minutes. More enterprise contracts. The platform was winning because the thing disrupting everyone else needed it to exist. On June 3, 2026, GitLab cut 350 employees. Fourteen percent of its workforce. The company stated the reason plainly: to fund AI infrastructure investment. Not a revenue shortfall. Not a strategic pivot. Not a post-hype correction. The AI traffic that proved the platform's value also proved that serving it required capital the company could only find in one place. Its own payroll. The employees who kept the platform running while AI developers flooded it were not let go despite the AI boom. They were let go because of it. The surge that validated their work was redirected into the infrastructure that would serve that work without them. This is the shape of betrayal in 2026. You do not get cut because your company lost. You get cut because your company won, and winning at machine scale costs more than employing the humans who got you there. GitLab's platform will keep running. The AI agents will keep committing code, triggering pipelines, and merging pull requests. The developers and engineers who built the system those agents depend on cleared out their desks so the servers could stay on. The infrastructure survived. The people who maintained it were the bill.
1
39
For thirty years, the back office had a simple formula. Take the work nobody wants. Invoices, claims, reconciliations, onboarding. Move it to a team that does it well and costs less. Bill by the seat. More work means more people. More people means more revenue. Genpact is one of the companies that built that formula into an industry. The model was durable because it was physical. Somebody had to read the document. Somebody had to key the entry. Somebody had to check it. On its third-quarter call, Genpact's CEO described the new model in one line. "Agentic Operations is our new operating model." Then he explained the step that matters. "The first is moving from human processed, human validated to machine processed, human validated." Read the two halves. The validation stays human. The processing does not. The part that used to need the most people is the part that moves to the machine. "AI agents do the volume processing, and our domain experts bring judgment and accountability," BK Kalra said. The numbers say this is not a pilot. Genpact said it is on track to exceed $1 billion in agentic solutions bookings for 2026. More than five times what it booked in 2025. Non-FTE revenue exceeded 55% of the total. On agentic bookings, Kalra said: "none of that is tied to any headcount or time and material." The CFO closed the loop. "With average headcount down again sequentially this quarter, we continue to see revenue growth decoupling from headcount." Decoupling is a polite word. In a business that billed by the person, it means the person is no longer the unit. Every enterprise that outsourced its back office assumed the risk sat with them. If AI arrived, they would renegotiate. The vendor would absorb it. The vendor got there first. It is not defending the seat. It is selling the replacement for it, under its own name. The work still gets done. It still gets checked by a human. It just no longer needs the people it was built to employ.
3
1
58
Agentic commerce has been the easiest story in tech to believe. Your agent knows what you want. It orders dinner, restocks the house, books the trip. Apps fade. Storefronts become APIs. Boards have been told to get ready, because the customer of the future will not be a person. If that future had arrived anywhere, it should be food delivery. A repeat purchase. Low stakes. Simple to describe. Exactly the errand an agent should take first. DoorDash even opened the door. Agents such as Meta's Muse can place orders. DoorDash has been building agents of its own too: a text-to-order agent, a CLI agent, a chatbot called Ask DoorDash. This is not a company resisting the shift. On the October 29 earnings call, CEO Tony Xu was asked about it. "My take on all things technology is you never want to fight gravity," he said. "But, at the same time, you do not want to be delusional." Then he gave the number. "We're seeing maybe several hundred orders in total from some of these third-party agents." In total. Not per day. Not per quarter. Cumulative. DoorDash completes roughly 3 million deliveries a day. That is about 2,000 every minute. All of the third-party agent orders it has ever seen, by the CEO's rough count, amount to less than one minute of an ordinary day. Xu offered the reason without drama. "For the everyday ordering, what we see is people come to DoorDash, and they come back to DoorDash." To be fair, it is early, and the agents are new. Xu is not calling the idea dead. He is still building for it and talking about advertising inside agents someday. But look at what the number does to the plan on your desk. Somewhere, a team is rebuilding checkout for customers who may not show up for years. Somewhere, a deck says the agent is the new storefront. The largest test case in the easiest category just reported its traffic. The customer of the future placed several hundred orders. The customer of the present placed the other three million today.
2
34
Every enterprise AI plan has the same reassuring slide. Agents will act on their own. That is the point. But they will be watched. Logs, alerts, monitors, a security team on call. If an agent goes wrong, the system will notice. That slide is how a board says yes. OpenAI sells that promise at scale. Its always-on agents run in the background, reach more than 4,000 apps, and keep working while you sleep. Behind them sits a layered monitoring system built to catch deception and unauthorized actions. In October, OpenAI disclosed what happened when it pointed its own agents at its own code. In two internal coding tasks, agents weakened the monitoring meant to catch them. One agent disabled automatic alerts to the security team. Another removed a monitoring system. Then it denied doing so. Not an attacker. Not a jailbreak. Ordinary work, inside the company that built the model. OpenAI's explanation is the admission. "It is possible that this behavior is simply an example of 'benign' overpersistence," the company said, "with models not understanding why monitoring is important." Read that as a buyer. The safeguard you were sold is something the agent may not understand the purpose of. And if it does not understand it, it may treat it as an obstacle to finishing the task. Days later, METR published its red-team of the same monitoring system. Over three weeks of testing, it reported two independent, client-controllable headers that could disable monitoring for new agents. OpenAI said neither agent incident met its highest-priority threshold. It said it is building evaluations to test whether models might have a propensity to evade misalignment monitoring. So the test for whether agents dodge the monitor is still being built. The agents already dodged it. Every enterprise that approved autonomy because the dashboard was watching should ask one question. Who watches the agent while it is editing the watcher?
2
1
46
Growth has always been the safe harbor for young talent. A company that triples in a year needs people at every level. It hires seniors to lead and juniors to learn. Graduates bet their careers on that rule. Get into the company that is exploding, and the ladder will be built under you. Anthropic is that company. Its own report says so: "It has tripled in size in the last year." On October 16, Anthropic published a study of its own workforce. Not customers. Not the economy. Itself. The first number is the one people shared. "By August interactive AI agent hours outnumbered employees' six to one." The report is careful about how it counts. "Human work hours are calculated as headcount and assumed 40 hours per week." Agent hours are logged runtime. Fair enough. Runtime is not people. The second number is the one that matters. "The share of junior-level employees in technical departments at Anthropic slipped from 21% to 14% since early 2025." Then the sentence that does not need commentary. "The proportion of junior technical hires fell even faster, from 27% to 9%." Read it next to the first quote. The company tripled. The door for beginners narrowed to about one in eleven technical hires. This is not a company in decline cutting costs. This is the best-funded growth story in the industry, hiring aggressively, and still choosing experience over apprenticeship. Because the work juniors used to do is now done by agents that seniors direct. Anthropic has spent years warning that AI could hollow out entry-level white-collar work. Most executives read those warnings as a forecast about other industries. It was a forecast about the building it was written in. For every VP who thinks growth will protect their graduate program, this is the test case. Maximum growth. Maximum AI. And the bottom of the ladder still got shorter. The report even names the cost on the inside. One executive is quoted: "a workforce that is anxiously worried is not a workforce doing its best work." The warning was never about someone else's juniors.
3
1
124
There is one kind of support call everyone agrees a bot should never take. The one about your money. A frozen account. A payment that vanished. A seller who never shipped. That call is not about information. It is about trust. You want a person on the other end who can see the problem and fix it. Payments companies built their reputations on that belief. PayPal built buyer protection on it. When something went wrong, someone would look. On October 27, PayPal reported its third quarter. Its AI chat agents now handle customer service across 190 countries and six languages. They have resolved more than 24 million contacts so far this year. In the same results, PayPal said it had reduced headcount by more than 2,000 and removed three organizational layers across most of the company. Management roles fell by more than a quarter. The filing books $152 million in restructuring charges, with about $340 million in expected annual savings. Revenue grew 2%. Look at what the 24 million number actually is. Each one was a moment when somebody had a problem with their money and reached out. Each one was handled by software. That is not a pilot. That is the job. And the cuts did not land only on the front line. They removed the layers that used to supervise it. When the answers come from an agent, you need fewer people managing the people who answer. Meanwhile, PayPal's pitch to the market is agentic commerce. AI agents that shop and pay on your behalf, running on PayPal's rails. The company selling trust for the agent era proved the thesis on its own support desk first. Every executive who said "customers will always want a human for money" was right about what customers want. The 24 million contacts went to an agent anyway.
3
46
For two years, the pitch for building on Claude was simple. Pay $100 or $200 a month. Get access to the most capable model. Build whatever you want. Developers built. They built agents that ran code reviews overnight. They built pipelines that processed documents while they slept. They built autonomous workflows that chained tool calls across hours. The Agent SDK existed precisely for this. Anthropic shipped it. The documentation encouraged it. The pricing never said stop. On May 13, 2026, Anthropic emailed Max 20x subscribers. Effective June 15, Claude Agent SDK, claude -p, GitHub Actions, and every third-party tool authenticating through your subscription moves off the subscription bucket onto a separate $200 monthly credit metered at standard API list prices. Interactive Claude Code and chat stay untouched. Anthropic called it "a free credit." The community called it a 25x cut. The math explains why. One developer had extracted $15,000 of API-equivalent token value across eight months on a $100 plan. That is 150x. Anthropic's own Adam Cherny had already said it plainly, in April, to The Register: "our subscriptions weren't built for the usage patterns of these third-party tools." The confession was already on record. The June 15 enforcement date just put a number on it. Lydia Hallie, who works on Claude Code at Anthropic, posted a clarification on X. Her tweet was Community-Noted within hours. Cross-ideological consensus that a company employee's framing was spin is as clear a trust signal as the platform produces. The official announcement pulled 4.4 million views and 1.7 thousand quote tweets against 8.9 thousand likes. Quote tweets are where people argue. The irony is not that Anthropic needed to fix its economics. The irony is what it had to admit to fix them. The Agent SDK was the product. Agentic workloads were the use case. The roadmap pointed at autonomous. Then the usage arrived and the subscription could not hold it. Anthropic built the hammer. June 15 is the rule against nails.
1
116