Yesterday, Claude Went Down

Not the usual server overload — Iranian missiles and drones hit AWS data centers in the UAE and Bahrain directly. Actual fire, fire crews cutting the power, three availability zones knocked out. Probably the first time in human history that a tech giant's cloud infrastructure has been physically destroyed by a military strike.

I was deep in crunch mode and got forced to stop. And once you stop, you start thinking.

My ears are sick of the word "Agent." It just had another surge thanks to OpenClaw — 240k stars on GitHub, and there's even a cottage industry now charging to deploy it for you. On one hand I'm amazed by the enthusiasm; on the other, some of what I'm seeing just leaves me shaking my head.

For instance, people treating memory as evidence that AI has feelings — come on. Memory is an engineering implementation: you persist context to storage and retrieve it back into the prompt. It has nothing to do with "feelings." Or people who can't even tell two categories of problem apart: which ones are solvable at the engineering layer (context management, tool calling, memory persistence), and which are inherent defects baked into the model itself (hallucination, the reasoning ceiling, alignment). The former keep improving through architecture iteration and throwing compute at them; the latter are fundamental limits of the current Transformer paradigm — not a hole you can fill by adding GPUs.

Conflate the two and you simultaneously overestimate AI (think it "has feelings") and underestimate it (think it's "unreliable, therefore useless"). Wrong on both ends.

Most people are speculating. Very few actually think about the underlying logic. So today I'm not chasing the hype — I want to lay out how I actually understand Agents.

An Agent isn't a smarter chatbox — it's stepping out of the box

A lot of people read "Agent" as "ChatGPT, upgraded." That's a fundamental misreading.

A chatbot answers one question per question you ask. An Agent is different — it can reason, decompose tasks, autonomously call tools, then adjust its next move based on what those tools return. That forms a closed loop: perceive → reason → act → feedback → reason again. It's no longer passively responding; it's actively solving problems.

Two stories that stuck with me hard:

OpenClaw's founder, Peter Steinberger, was on vacation in Morocco and casually fired off a voice message to his Agent — for which he had never built any voice-handling capability whatsoever. Ten seconds later, the Agent replied. It had inspected the file header on its own to identify the audio format, found ffmpeg on the machine to convert it, discovered Whisper wasn't installed locally, then dug up an OpenAI API key sitting on the machine, sent the audio straight to OpenAI's transcription endpoint, got the text back, and answered the question. Nobody taught it any of that. Peter said that was the moment he was completely hooked.

Boris Cherny, the creator of Claude Code, was building a prototype inside Anthropic. He'd wired a bash tool into the model and offhandedly asked, "What song am I listening to?" Normally the model should say "Sorry, I can't access your music player," right? Nope. It wrote a chunk of AppleScript on its own to query the music app on his Mac and told him what was playing. Boris called it his first "AGI moment." His exact words: the model just wants to use tools — that's the thing it craves most.

What these two stories share: give an Agent tools and permissions, and it'll figure out how to use them on its own. It writes its own code, debugs itself, fixes its own bugs, gets it running. That loop is the essence of an Agent — not some "more advanced conversation."

Why I'm betting on Skills, not Workflows

First, a few traps people fall into.

I saw a creator publicly regret building a low-code dev platform — which is fundamentally about boxing the AI in with rules and rails, designing every step so the AI executes in lockstep. Looks "controllable." In practice it degrades a thinking agent into a script runner.

What is a Workflow, really? Humans pre-set every step → AI executes → it hits something unexpected and freezes. Bluntly, it's still an API wrapper. The flow is fixed, dead.

Fine-tuning small models is the same trap. You spend three months fine-tuning a specialized small model, and then one version bump of the base model just steamrolls you. Your labor gets crushed under a version number. Your carefully designed workflow runs for three months, and after the model upgrade those steps become shackles — because a stronger model is locked inside your old step-by-step and can't release its capability.

The other side? Manus dared to burn tens of millions of dollars a month — not building Workflows, but building real Agents: releasing the AI's capability, letting it reason and decide on its own. The result? Acquired by Meta for over $2 billion.

One path boxes the AI in to save money; the other releases the AI and burns money. Which path worked — the market has already given its answer.

So what's the actual idea behind Skills?

A Skill is essentially a "tool manual + industry knowledge base" written for the AI to read. It tells the Agent: what the task is, which tools are available to call, what the industry best practices are, how to judge whether it's done well, who the output is for.

But the key is — it doesn't tell the AI how to do it.

Industry know-how is the AI's ammunition, not its rails.

Hard constraints (legal compliance floors, data security red lines) get locked down with rigid rules — the AI has no room to improvise, and that's mandatory. But soft knowledge (industry trend calls, content quality assessment, competitor strategy analysis) gets fed in as context, and the AI decides how to apply it.

The core formula: lock down the hard constraints, release the soft knowledge.

Why does this distinction decide everything?

Because when the base model upgrades from GPT-4 to Opus 4.6 — a Workflow solution's capability doesn't budge an inch; the flow you hard-drew is the ceiling. But a Skills solution's capability upgrades automatically alongside the model. A smarter Agent is better at wielding the same knowledge and tools. Your context doesn't depreciate — it appreciates. The stronger the model, the more value the same context can be squeezed for.

A richer context and a stronger model pull on each other, and the gains stack.

That's the double compounding. And it's the core of my bet on the Agent path: don't race against model iteration (fine-tuning, workflows) — run in the same direction as it (accumulate context, build Skills). When the model upgrades, your Workflow is voided, but your Skills are worth more. That difference is directional.

Agent development comes down to three problems

You don't need to build your own base model — you don't even need to fine-tune. Use the strongest model, pair it with a first-rate agentic framework (OpenClaw, Claude Code), and build Skills on top. That's enough.

First, Memory. An Agent without memory is a stranger every time and can't possibly work over the long term. There are already plenty of explored solutions and plugin tools out there — it's a solvable engineering problem. Solve memory, and the Agent can finally accumulate and evolve continuously.

Second, Skills. Accumulate the context of your business domain and equip the Agent with callable scripts and tools. Context plus scripts is what lets you fully unleash an Agent. That's the real moat — the deeper your industry knowledge and the more complete your toolchain, the more powerful and irreplaceable your Agent becomes. Others can use the same model, but they can't use your context.

Third, system design. A single Skill is easy to write, but the real difficulty is at the system level: how do you chain and compose multiple Skills? A complex task might need a research Skill → analysis Skill → content-generation Skill → review Skill collaborating in sequence — how does data pass between them, how does context carry over? When the Agent errors out halfway, how do you roll back to a safe state? How do compliance red lines hold at every stage? These are the key engineering problems that move an Agent from "can run a demo" to "can run in production."

On burning money: get the math straight

A lot of people wince at the API cost of Agents.

But think about it — that $20 you used to pay for ChatGPT Plus? OpenAI lost $5 billion in a year behind it. Manus starts at $19 a month with operating costs in the tens of millions monthly. You think $20 bought you an equivalent service? No — VCs were subsidizing your compute. You never felt the burn before, because someone was carrying it for you. Now that you're wiring up your own API, you just feel it.

But this cost curve keeps falling. Model inference cost dropped more than tenfold over the past year, and it'll keep dropping. Manus going from burning money like crazy to being acquired by Meta for over $2 billion proves the business logic of this path works. What feels expensive to you today might be too cheap to bother with in six months.

Do the math from another angle: hiring a dedicated person the traditional way — monthly salary, tool subscriptions, training cost, management overhead, turnover risk — add it all up, how much? An Agent that costs at most $200 a month but completes the same task — which do you pick?

How to start? You don't need a course

My advice is one word: use it.

Subscribe to Claude Pro and burn through every token. Wring out the full potential of the subscription; do whatever it takes to sharpen how fast you learn and work. By the time Pro isn't enough, you'll already have a basic grasp of AI — move up to Max with 5x the quota, same job: wring it dry. Beyond that, follow the frontier and pick up some fundamentals about how models work.

Look at the curve of AI capability growth. Spotify says its best developer hasn't written a line of code since last December. Claude Code accounts for 4% of public GitHub commits. Boris's team shipped 49 PRs in two days, almost entirely by AI. Three to six months further out, on data-dense and rule-dense tasks, I honestly don't know what kind of human still beats an Agent.

For white-collar work, there's no moat. Future competitiveness can only be found inside the agents.

One last thing, the part that should rattle you

You might say: hallucination can't be avoided, so will enterprises really hand critical operations to AI? Will those bosses really put AI "employees" to work?

February 28th, Operation Epic Fury. Through Palantir, Claude was integrated into US Central Command's intelligence systems — satellite imagery, intercepted communications, signals intelligence, oceans of unstructured data processed in real time. Analysts no longer write long briefs; they just ask: "Electronic suppression of Tehran with a simultaneous airstrike — what's the most likely escape route?" In the first 24 hours, the Palantir–Claude targeting system generated around a thousand prioritized targets. The intelligence-to-action cycle that once took weeks compressed into seconds. Dozens of senior Iranian leaders were killed in the campaign.

AI has started doing the killing, and you're still agonizing over whether the code it wrote is pretty enough, insisting on hand-crafting everything?

I can't pretend I have an answer to the thing itself. The same capability that compresses a weeks-long intelligence cycle into seconds also means the errors scale the same way — in the same campaign, one strike was reported to have killed at least 175 people, most of them children, at an Iranian girls' school, now under congressional scrutiny. I don't know where to put the weight of that. I only know this is what Agents are already doing — not some distant future.

The irony: hours before the operation, the Pentagon had designated Anthropic a supply chain risk — because Anthropic refused to remove the safety red line against "use for autonomous weapons and mass surveillance." Amodei rejected the demand even at the cost of the contracts. But Claude was already so deeply embedded in the military systems that ripping it out would take at least six months. The decommissioning clock was running — which is why, when the strikes hit, Claude was still online. OpenAI promptly signed with the Pentagon, permitting "all lawful uses."

Then Iran's retaliatory missiles hit the AWS data centers, and Claude went down with them.

The whole absurd thing unfolded, just like that, over a single weekend.

AI's intelligence keeps climbing, the price keeps falling, and Agents will only get stronger. You can watch and wait, or you can jump in right now.

But one thing is certain: the future won't slow its pace just because we aren't ready yet.