Anthropic engineer: "You're not supposed to babysit the model. Put it in a graph and it catches its own mistakes, running a dozen tasks at once." In 25 minutes he breaks down exactly how Anthropic builds agents that run in parallel, check each other's work, and recover when one... fails. Worth more than any paid course you'll find on building agents. Watch it, then read detailed guide on graph engineering below.show more

not supposed to babysit the model” is the whole thing but the babysitting doesn’t disappear, it moves upstream. if step 1 is vague, four workers produce four flavors of vague and merge averages them into mush most of my time goes into the goal node now, not the prompts

@AnatoliKopadze that's wild, didn't know you could make them self-check like that. need to rethink how I handle error correction in my projects.

i'll watch this video and i'm sure i'll learn something new, thanks for the info

This is literally somebody else’s science, which do you like to figure out. This is somebody’s biometrics that they’re taking the concepts from.

lol no it builds slop

Parallelism helps most when the tasks have different failure modes. Giving every agent the same job faster just creates a larger pile of correlated mistakes

AIを管理するより、AI同士が確認し合う方が人間らしい働き方になるのかもしれませんね🤔

Better models aren’t the whole answer. Sometimes the biggest performance gain comes from better orchestration: split the task, run work in parallel, verify outputs and recover from failures. Architecture matters. 👀

parallel self checks still share one graph control plane risk

Catches its own mistakes" is doing heroic work there.

“Put it in a graph” isn’t the hard part. Owning the graph is. Parallel agents that check each other still need a home for state, a stop rule, and a human who stands behind the blast radius — otherwise you’ve just automated babysitting at higher token cost.

The part that actually matters is the error recovery design. Most agent demos show happy paths. The interesting question is what the graph does when two parallel nodes return contradictory outputs, does it arbitrate, escalate, or just pick one silently?

honestly the parallel checking thing is such a game changer for reliability

With opus 5, yes - you do have to babysit. And instead of getting payed to babysit you owe THEM money and it ruins your day everytime.

Right. Another concept stolen from its users. Good luck with getting no regression s!

agents checking agents feels right

The graph moves the babysitting upstream. That part is clear. What still feels open: when a node fails and the graph recovers, who owns the corrected state afterwards? The graph can catch mistakes. It still needs a durable layer that outlives any single agent.

Parallel agents that self-correct feels like the first time these systems stop feeling like toys. Anyone actually running production graphs yet and seeing the failure modes?

Absolutely BS

This parallel processing approach highlights how robust systems emerge from distributed checks, a core concept for translating consumer attention into investor foresight.

“Put it in a graph and it catches its own mistakes” is the sentence that separates people who build agents from people who babysit them.

Don’t you think all “right” approaches proposed by model creator coincidentally are also tokenmaxing?

The interesting shift is from babysitting agents to designing systems that supervise themselves.