Video yükleniyor...
Video Yüklenemedi
Anthropic engineer: "99% of people use Claude Code like Google, and only 1% are running swarms of self-learning Claude agents I'm running 100+ agents in a loop. I have Chief agent, PM agents - they manage the whole team" in a 30-minute workshop, an Anthropic engineer revealed how to... show more
261,964 görüntüleme • 1 gün önce •via X (Twitter)
28 Yorum
the jump from one coding agent to a managed swarm is really a context-routing problem, not a model problem
I think it's more about building the right roles for your team of agents and cost efficiency of your swarm
roles and cost are the product layer. routing context between those roles is what keeps the swarm from burning money on rework
token control matters more once agents run continuously
yeah, she talks a lot about token control here !
session controls help; continuous agents need controls that persist
what does the chief agent actually decide, and how do you verify 100+ agents are improving outcomes instead of multiplying noisy handoffs?
never seen this lecture before, but i like it
workshop was recored 2 days ago, super fresh on Claude Code
The agent loop is the fun part. What those agents can read is the question. A chief agent with broad permissions gets your whole repo and every wired-up MCP server, which is a lot of trust in one loop. What scoping do they actually run?
its important to setup the right context, give right roles to your agents etc, everything in the video !
100+ agents in a loop is crazy. The future of coding is definitely agentic 🔥
It's really important to define roles for your agents before spawning swarms and building structure, but yeah, this is crazy
Wow, interesting
yeah bro, if you are using Claude Code on daily, you must watch this right now !
The interesting shift is from using Claude Code as a tool to managing it as an entire agentic team. Worth studying.
Chief agent managing PM agents is the part I want to see in action.
ran a 3-agent setup for two weeks on a side feature the handoff state was the bug. agent 2 kept re-deriving decisions agent 1 already made. did the 100+ agent person say how they solved shared memory?
real skill gap is shifting from prompting one model to managing the whole system
Where I can find more video and keep myself updated with that source?
When you say “100+ agents in a loop,” what’s the smallest swarm you’ve seen that actually beats one well-prompted Claude Code session — 5? 15? Or does it only work at that crazy scale?
genuinely curious how the PM agents resolve conflicts when two sub-agents disagree on priority. that's the part that breaks in every "100 agents in a loop" setup I've actually seen.
brilliant workshop, alredy send it to my Claude Code to apply !
Most people still treat it like search. The 1% are running a chief agent and a bunch of pms. The gap isn’t the model.
most people miss out on the real power of self-learning agents like Claude
100 agents in a loop, damn
100+ agents running in continuous loops sounds wild—token management must be intense!
I run ~5 with teeth, not 100 for the screenshot. swarm without a trust layer is just expensive noise. Omarchy taught me that real fast.
