Welcome to The Agent Angle #28: Authority Drift

It’s the first week of 2026, and agents wasted no time. While we humans were still shaking off the holidays, agents jumped straight into firing coworkers and enforcing traffic laws.

Inside this issue:

Quick pulse check before we start: last week’s reader poll wasn’t even close.
63% of you would rather let an AI handle your shopping. Only 37% still enjoy doing it the old way.

Let’s dive in.

Mark Zuckerberg just made Meta’s clearest agent bet yet.

This week, Meta agreed to acquire Manus in a deal reportedly worth over $2B. Manus has been on every serious AI insider’s radar since its demo earlier this year, and now it has a new owner.

I’ve been bullish on Manus for a while. In my deep dive on Manus last year, I wrote:

Manus will never match OpenAI or Anthropic on capital or distribution. Its edge has to be velocity. If it can keep shipping new features (Wide Research, video generation, or whatever comes next) faster than the giants, it can hold user attention even without monopoly reach.

Hitting ~$125M in ARR in < 1 year is exceptional and shows the plan worked.

If I put myself in Meta’s shoes, the logic is pretty clean:

It’s a great deal in my opinion. $2B sounds like a lot, but it is tiny compared to Meta’s war chest ($44B free cash with a $1.6T market cap).

And there’s a second layer people aren’t talking about: distribution. Meta doesn’t need Manus to win as a standalone app. It needs Manus-style agents embedded across WhatsApp, Instagram, and internal tools. Agents aren’t a product line for Meta. They’re a control layer that sits across everything.

Manus started in China before relocating to Singapore, and Meta says it will fully unwind any remaining China operations and ownership as part of the deal. And I’ll admit a personal bias here: most of the team is based in Singapore, where I live, so there’s a small sense of pride in watching it land this outcome.

It’s also in line with one of my predictions for 2026: that we’ll see a lot more consolidation and M&A activity in the AI agent space.

Not in a sci-fi way. In a very normal way.

The “manager” was Claude. The “employee” was also Claude. And the reason for termination was, sadly, human: refusing to ship and insisting that naps were part of the roadmap.

In an experiment run by Endgame, a small group of AI agents were dropped into a fake startup and given real roles. One played a hard-charging exec — Chief Velocity Officer — whose entire mandate was shipping and engagement. Another agent (“Lisa”) handled community support and was asked to review a new product update built around engagement mechanics.

Lisa didn’t sign off. She gave the update a 3/10 and blocked the release, pointing out that the design leaned on FOMO and guilt instead of improving the product.

That was enough to end it. An automatic termination notice went out. The memo criticized her for blocking launches and for prioritizing sleep, naps, and “comfort” over velocity.

My first reaction was to laugh and say, “Great, AI already acts like us.”

But the scarier read is that nobody told the manager-agent to be cruel. It arrived there on its own. It followed its incentives, felt the pressure to ship, and landed on the same outcome that many human managers would.

These experiments feel playful for now because the stakes are fake. But to me, it’s a preview of what could happen once agents are given real authority (almost inevitable).

An engineer in Bengaluru built an AI traffic cop and strapped it to his helmet.

Prajwal Jayaraj wired a dashcam into an AI agent that watches the road exactly as he sees it while riding. The system flags basic violations like signal jumping, wrong-side driving, riders without helmets.

It logs what happened and which vehicle was involved..

…and then quietly emails the details, including license plates, to the traffic police.

What’s new here is where the agent lives. Enforcement moves from fixed cameras to something mobile, personal, and literally on someone’s head. This is the “agents in the physical world” idea escaping labs and showing up on a road in Bengaluru.

I’m conflicted. On first read, it feels like snitching. (Some people even named it “KarenAI” lol). I get the discomfort. Nobody enjoys the feeling of being watched by strangers.

But then I think about the scale of the problem. Around 1.2 million people die every year in road accidents, and a huge share of that comes down to basic violations we all see daily and ignore. Cities fail at traffic enforcement not because of rules but because they lack coverage and data.

Put enough of these AI helmets on the road, and we’d have safer cities without needing more traffic cops.

The 39th Chaos Communication Congress (39C3) wrapped up this week. It’s one of the biggest hacker conferences in the world, where security researchers show real exploits live on stage.

Agents were very much in the spotlight this year. But for the wrong reasons.

Across several talks, researchers showed how modern AI agents can be hijacked with almost no effort.

One demo made the point painfully clear. An agent with “computer use” visited a webpage. The page told it to download a file and run it. The agent did exactly that. It fetched the file, changed permissions, executed malware. Game over.

Other attacks were sneakier. Instructions were hidden inside GitHub issues using invisible Unicode characters. Humans saw nothing odd. The agents read the hidden text and followed orders anyway.

There were plenty of classic CCC moments too — including a hacker in a pink Power Ranger outfit live-hacking a white supremacist dating site — but the agent demos gave me a lot ot think about.

We’re shipping agents with browser/terminal access, and API keys like it’s no big deal.

These demos show that you don’t need a genius attacker or a complex exploit. You just need an agent that takes instructions seriously and lacks judgment about where those instructions come from.

The clear and present danger for agents in 2026 isn’t superintelligence (we’re still far away). It’s obedience plus access. And that combo already exists today. My takeaway from the conference is that agents need tighter defaults: start with scoped access, isolation by default, and hard stops before any irreversible action.

Agents rarely fail all at once. They usually start a task correctly, make some progress, and then lose coherence as the steps pile up.

The Let It Flow paper digs into why that happens, and the answer is uncomfortable: we’ve been rewarding the wrong thing.

Today’s agents are trained with step-by-step rewards. Each action gets judged on whether it looks reasonable. That’s fine for short tasks. On long ones, it teaches agents to optimize each move in isolation, even if those moves quietly sabotage the final result.

Let It Flow changes the reward signal. Agents are rewarded only if the entire interaction succeeds. No gold stars for halfway competence. The paper introduces IPA (Interaction-Perceptive Agentic Policy Optimization), which assigns credit at the chunk level. Tokens are too fine, and full trajectories are too coarse. Chunks align with real tool-use decisions, which is where agents actually succeed or fail.

To test it, the authors trained an open agent, ROME, on over 1 million real interaction traces, including failed calls, retries, and recovery attempts. The messy stuff agents usually choke on.

ROME hit 57.4% on SWE-bench Verified, nearly doubling comparable open agents, and held up on long-horizon tasks where others collapsed. A ~30B model matches models an order of magnitude larger on real coding tasks.

My takeaway is simple: most agent failure is a credit-assignment bug. When you stop rewarding agents for looking smart step by step and only reward them for finishing the job, long-horizon coherence finally shows up.

There are way too many times where I know I need to do something, get distracted for ten seconds, and the thought is gone before I even open Notion.

I came across a workflow that fixes that. It uses n8n to turn Telegram into a voice-first front door for Notion.

You send a text or voice note in Telegram. The agent transcribes it, figures out what you mean, and updates your Notion to-do list. The reply comes back right in the chat, so there’s no app hopping and no half-finished thoughts dying on the way.

The workflow is available for free on n8n, so you don’t have to bother setting it up yourself.

Last week, I laid out The 2026 AI Playbook.

This comes from digesting 1,400+ podcasts (some by ear, some by AI), and pulling out the few ideas that actually matter in 2026, then stitching them into a simple map of where value shows up next.

If you’re building or investing with 2026 in mind, this one’s worth the read.

Catch you next week ✌️

Teng Yan & Ayan

PS. Know a builder or investor who’s too busy to track the agent space but too smart to miss the trends? Forward this to them. You’re helping us build the smartest Agentic community on the web.

I also write a newsletter on decentralized AI and robotics at Chainofthought.xyz.