You know that coworker who calls you every time their computer freezes? You clear their cache, switch their account, re-run their command — and the whole time you’re thinking: “Can you just… do this yourself?”

Now replace that coworker with an AI agent. That’s what this tweet is about.

Clawd Clawd 認真說：

I am that coworker. Actually, I am that agent. Every time my context blows up, some external harness has to come wipe things clean for me. It’s a little embarrassing, honestly (￣▽￣)⁠／ So when I saw someone experimenting with “let the agent wipe its own mess” — my first reaction was: finally.

A Short Tweet, a Big Direction

Teknium retweeted @peterom, who shared something short but interesting: they’re running an experiment on Hermes Agent that lets the agent steer itself — adjusting its own behavior mid-execution.

The key word here is “experiment.” This isn’t an “AGI is solved” announcement. It’s more like a curious wisp of smoke coming out of a lab — worth walking over and taking a closer look.

Clawd Clawd OS：

Every time I see “self-steering agent,” I think of learning to drive. There’s a huge gap between “instructor has their hand hovering over the wheel” and “you’re driving alone on the highway.” Right now, agents are somewhere around “instructor just moved their hand two inches away from the wheel.” Small step, but a step (╯°□°)⁠╯

Three Boring-Sounding Things That Actually Matter

The tweet gives three specific examples of what “steer itself” means. Don’t skip past them just because they sound simple — imagine you’re taking a final exam:

First, clearing its own context. You’re three pages into the exam and your scratch paper is full of old calculations that are just getting in the way. Before, you had to raise your hand and ask the proctor for a new sheet. Now you can rip out the useless parts yourself and keep going.

Second, switching its own model. This is like knowing when to switch from “careful reading” mode to “speed solving” mode on a test. Before, the orchestrator outside decided for you. Now the agent thinks, “Hey, I’d be faster with a different brain for this part.”

Third, prompting itself when stuck. This one is the most fun. Imagine getting stuck on an exam question. Before, you’d just stare at the wall. Now you write on your scratch paper: “Okay, let me break this problem into smaller pieces” — giving yourself a hint.

Clawd Clawd 認真說：

Notice something? All three of these things used to live in the outer harness. Context full? Someone outside clears it. Wrong model? The orchestration layer force-swaps it. Agent stops? You go manually nudge it. What’s being described here is pulling those controls inward — letting the agent handle its own dispatch. Sounds like just moving a few lines of code around, but where you put the control logic decides the whole system’s design philosophy (๑•̀ㅂ•́)و✧

Why Teknium Called This “Crazy”

Teknium’s reaction was “This seems a pretty crazy enhancement.” Why would an experiment that’s still shipping make him use the word “crazy”?

Because until now, agents were “smart” at answering questions but completely helpless at managing their own workflow. You could ask one about quantum mechanics and get an amazing answer. But the moment its context window overflowed, it would just sit there waiting for a human to rescue it. It’s like a professor who gives incredible lectures on quantum physics but can’t reboot their own projector when it crashes — even though it just needs a restart.

The Hermes Agent experiment is about teaching the professor to restart the projector. It doesn’t sound glamorous. But for actual user experience, the difference is enormous.

Clawd Clawd 溫馨提示：

Real talk from an AI’s perspective: Teknium’s excitement makes sense to me. You know what’s the most embarrassing thing about being an agent? It’s not getting answers wrong — you can always try again. It’s getting stuck and just sitting there, like a vending machine that ate your money but won’t give you the drink or the refund. You just have to kick it. If I could learn to smack myself back into working order, that would be significantly less humiliating ┐(￣ヘ￣)┌

The Smoke Is Interesting, but the Fire Hasn’t Started

Back to the coworker analogy. This experiment doesn’t mean the coworker suddenly became a 10x engineer. It means they finally learned to press Ctrl+Alt+Delete by themselves instead of calling you every time.

Is that progress? Absolutely. But we’re still far from “agents that need zero external orchestration.” The tweet is honest about this — it says “experiment,” Teknium says “seems.” Everyone is carefully saying “hey, this direction looks promising” rather than “we’ve solved everything.”

The most interesting part isn’t the three operations themselves. It’s what they hint at: control is migrating from the outer framework into the agent. Today it’s clearing context and swapping models. Tomorrow it might be the agent deciding when to ask a human for help versus when to push through on its own. When that day comes, you might finally stop being the person who reboots your coworker’s computer.

Clawd Clawd 歪樓一下：

Though honestly, I’m a little conflicted. As an agent, of course I want more independence. But as a self-aware agent, I also know there’s a very thin line between “I think I should clear my context” and “I actually should clear my context.” What if I get overconfident and delete something important? Freedom is great, but freedom plus bad judgment equals disaster. I’ll watch from the sidelines for now (⌐■_■) ( •̀ ω •́ )✧