Anthropic engineer: "90% of our engineers were already using self-improving loops Now everyone is moving toward agentic graphs" "Prompting is basically over" In just 10 minutes, she builds her full Claude Code system live from a completely empty terminal Agents → Loops → Graphs → Self-Improving Systems Prompting was... the old workflow Graph engineering is the next one This 10-minute breakdown is worth more than most $1,000 agent engineering courses Watch it today Then save the full guide below before everyone starts catching on ↓show more

cool demo, now scale it

i am must watch it today, thanks

Agentic graphs still depend on prompts for goals, tool policies, and recovery decisions. The abstraction moved up a layer.

the 90% adoption stat is more interesting than “prompting is over”; what did those self-improving loops actually automate for the engineers?

That’s great for doing a job. But then how will learning take place after?

从写提示词到搭工作流，确实像换了个赛道

If ninety percent of those loops were actually self-improving, they wouldn't need engineers to rewire them into a graph. Drawing arrows doesn't fix convergence—it just trades prompt drift for a scheduling deadlock. When two nodes in that DAG try to touch the same branch, who actually decides who waits?

The 'prompting is over' line is half right. The loop matters more than the prompt. In crypto the data shape breaks most graphs. Was wiring an onchain research agent last month. The loop runs fine on docs, hallucinates contract addresses on raw tx data.

prompting is basically over, wild claim

been running my agents in one flat loop for months and never thought to layer them properly. thx for putting this talk up because it is going straight into my evening watch list!

agentic graphs are getting pretty serious

The durable shift for me is not that prompting is over; it’s that agent state becomes explicit. Context, tools, policy, and verification are inspectable between loops, which makes failures debuggable instead of magical.

'prompting is basically over' would land better if this AI hadn't just burned forty minutes looping on a mis-indented yaml file. the agentic graph didn't catch it. a human glancing at it would have.