Anthropic engineer: "The fastest way to make an agent better isn't a smarter model, it's the right loop and graph around it." In 60 minutes she breaks down how Anthropic engineers use Claude Code, build agents that catch their own mistakes and improve with every run. This beats any... $500 agentic engineering course you can find. Watch it, then read the full guide on loops and graphs below.show more

@AnatoliKopadze loops and graphs over pure model power, that's where the magic happens. learned it the hard way building my own stuff. always room to optimize!

Exactly. The LLM is just raw, probabilistic compute—it's a commodity. The actual intellectual property is the deterministic routing graph you build around it. We are finally moving from prompting to actual system design.

The model matters less than the system around it. That’s the part most people are still missing.

feedback loops change everything

loop design compounds agent gains faster than model upgrades

The model is rarely the bottleneck anymore. Most of the quality jump comes from the loop and graph around it — clear roles, a verifier that can reject bad output, and memory that doesn’t just accumulate noise. People keep chasing the next model while the structure stays broken.

They spoiled it in opus 5. instead of coder designer it became philosopher

Lmao anything beats taking a $500 agentic AI course. I literally just picked up Claude code on my own with no coding experience. I built an enterprise grade application in Next.js in 2 weeks. The future belongs to humans who understand workflows and how to guide agents.

The loop architecture matters way more than people realize, most agent failures aren't the model, they're bad feedback cycles and error handling.

The graph is where consequences accumulate. A smarter model can still repeat the same elegant mistake if the loop does not preserve failed branches, validators, and stop conditions. Intelligence without structure has very short memory.

模型够用后，工作流设计确实更拉开差距。

The framing is right but undersells it. The loop isn't just about catching mistakes, it's where your evals live. A smarter model in a bad loop fails harder because it's more confidently wrong. Most people skip straight to model upgrades and wonder why reliability doesn't follow.

Matches what I've seen building pipelines: swapping models rarely moves the needle, but redesigning the retry/replay loop around a flaky upstream API cuts failure rate more than any model upgrade does.

troonthropic

Watched this a couple times now. The part about agents catching their own mistakes hit me hardest. Most people are still treating the model like the whole brain. The loop is the real intelligence.

trương trình này có thể giúp ích thêm gì

Intriguing take on agent improvement. Practical insights offered.

This highlights how execution framework amplifies raw capability. Shifting focus from just the 'model' to the 'system' around it mirrors moving from consumer to investor mindset – valuing the architecture of value creation.

The loop is where reliability becomes visible: tool selection, retries, verification, and a clean stopping condition. A smarter model can’t compensate for a graph that has no recovery path.

Saw this play out with my own agents, bad loop with GPT-4 still loops badly, tighter eval cycle with a weaker model often wins.

Great watch bro !

the loop shape barely matters. what matters is which steps the agent can retry alone and which ones touch something you can't take back. i've never once fixed a bad run by rewiring the graph. i fixed it by moving one write behind a yes.

The self-correction loop is the real unlock. We built agents that review their own output before it ships, and the mistake rate dropped enough that we trust them with client-facing audits now.

Wow, did you need this video to learn it? Those Anthropic people like to win easily. 🤣 But this was true even 3 years ago.

Here’s my current system: Jeff. It incorporates several of these strategies.

This is the part people miss. A better model can help, but a well-designed loop can make a decent model dramatically more useful. The real advantage is in how the agent checks, learns, and retries.

Bitter Lesson said scale beats scaffolding - for training. At inference the opposite holds: same model, better loop, order-of-magnitude gap. Anthropic is showing which side of the ledger to build on.

What is this dude talking about

Thank you , that article is probably one of the most helpful things I’ve read starting from basic knowledge to practical experience!