1mpulse

Finding the AI ideas builders will use next.

Agents, models, workflows.

[1mpulse @0x1mpulse]

https://t.co/T151L051w2

[1mpulse @0x1mpulse]

@0xCodez Useful roadmap for one of the most interesting roles emerging in AI.

[1mpulse @0x1mpulse]

@de1lymoon Strong practical walkthrough. Connecting search, tracing and deployment makes it especially useful.

[1mpulse @0x1mpulse]

@0xMorlex Excellent breakdown. The four-layer model makes this much easier to understand.

[1mpulse @0x1mpulse]

Primary sources behind this guide:

Anthropic: Demystifying evals for AI agents
https://t.co/QyBmUT02wb

LangChain: Agent evaluation approaches
https://t.co/71l7o4vDI7

[1mpulse @0x1mpulse]

https://t.co/T151L051w2

[1mpulse @0x1mpulse]

@EXM7777 This is one of the most creative Claude Code workflows I’ve seen.

[1mpulse @0x1mpulse]

@de1lymoon Great find. The evaluation section looks especially useful.

[1mpulse @0x1mpulse]

@0xCodez Excellent resource. The timestamps make this especially useful.

[1mpulse @0x1mpulse]

@0xCarnagee The shift from one loop to specialized agents is the most useful takeaway here.

[1mpulse @0x1mpulse]

@0xVampas Love how clearly you turned this into an actionable playbook.

[1mpulse @0x1mpulse]

@0xMorlex Great breakdown, especially the focus on verification before production.

[1mpulse @0x1mpulse]

@carm1nee The real moat is turning expensive tokens into measurable business value.

[1mpulse @0x1mpulse]

@vartekxx Long context matters, but knowing what to exclude may matter just as much.

[1mpulse @0x1mpulse]

@choopyplug1 The recursive loop - AI systems improving AI systems - is the part worth watching most closely.

[1mpulse @0x1mpulse]

“Coding is solved” sounds like hype.

But Boris Cherny, the creator of Claude Code, adds an important qualifier: it is solved for his workflow.

Claude now writes 100% of his code. The harder problem is managing what happens around it:

00:31 - why coding is “100% solved” for him
01:00 - dozens of AI-written PRs per day
01:28 - where coding agents still struggle
02:28 - hundreds of agents across multiple sessions
02:58 - loops that keep working in the background

His agents fix CI, rebase pull requests, and organize user feedback continuously.

The bottleneck is no longer producing code.

It is designing the system that decides what should run, what should be trusted, and what still needs human judgment.

[1mpulse @0x1mpulse]

The biggest shift in coding wasn’t a new programming language.

Andrej Karpathy says his workflow flipped from roughly 80% writing code himself to 80% delegating it to agents.

The interesting part is what happened next:

00:29 - coding becomes delegation
02:33 - failures start looking like workflow problems
03:28 - agents receive independent “macro actions”
03:44 - research, planning, and implementation run in parallel
04:32 - unused tokens begin to feel like idle GPUs

The bottleneck is moving from typing code to deciding what multiple agents should do, keeping their work separate, and reviewing the result.

Agentic engineering is becoming a management problem disguised as a coding problem.

[1mpulse @0x1mpulse]

@0xCodila Better agents often need less context, not longer prompts.

[1mpulse @0x1mpulse]

@0xMortyx The scaling strategy matters more than another leaderboard win.

[1mpulse @0x1mpulse]

@tomasSAXC @claudeai Planning before touching the code is the real upgrade.

[1mpulse @0x1mpulse]

@HeyAnjula @JensenHuang Loops turn prompting into a repeatable system.

[1mpulse @0x1mpulse]

@0xMovez Nice find, Movez. Saved this one.