Kushal ᯅ

Aspiring Chief Claw Officer 🦞 Analytics Engineer by day 🧐 Building agents by night 🤖Tinkerer at heart! ❤️

[Kushal ᯅ @kushal_to]

Current state.

Really love the added window controls.

[Linghua Jin 🥥 🌴 @LinghuaJ]

Subagent or no subagent? I've been noticing this for a while. Some high level thoughts:

Traditional coding tasks are deterministic: clear input, clear output. You can hand them off. Agentic engineering is different. Most of the time, the boundary isn't clear.

Simple tasks have clean boundaries. Most real tasks are tangled. Example: a codebase-explorer subagent with its own context window. The parent gets a summary. Then I spin up a new subagent for a specific change, and it needs details the summary dropped.

Not everything can be isolated. Even in the main thread, during exploration you don't know what you'll need later. The "irrelevant" detail you skipped is the one the next step depends on.

It's the same as human teams. One person vs. a 100-person team: 100 people don't give you 100x. Communication is overhead. Every handoff loses something. Agents are no different.

So the real problem to solve isn't "how many agents." It's knowledge synchronization efficiency: how much context survives each handoff, and at what cost.

That's the bottleneck for multi-agent systems, same as it always was for teams.

[Kushal ᯅ @kushal_to]

@chase_mccoy @claudeai Why does this look so much worst

[Kushal ᯅ @kushal_to]

@chase_mccoy @claudeai Hmm

[Kushal ᯅ @kushal_to]

@aiwithadb @AnthropicAI @bcherny @lydiahallie @ClaudeDevs Key is likely to limit subagent usage.

[Kushal ᯅ @kushal_to]

@DavidSHolz Not everyone shares the same view unfortunately. Pun intended.

[Remi Fabre @RemiFabreRobot]

Sorry for the wait, the episode is here!

Live action with real robots instead of animation (episode 4): "Anger"

Enjoy!

As always, the software of Microduck and Reachy Mini is open source:
https://t.co/RW1p4V2X6X
https://t.co/hA91qMLZ4d

[Kushal ᯅ @kushal_to]

@RemiFabreRobot Omg. 😭😭😭

Reachy will walk one day! We will make it happen.

[Don Allen Stevenson III @DonAllenIII]

😮 it’s even better in a spatial computer

[Kushal ᯅ @kushal_to]

@DonAllenIII I love this so much.

[Kushal ᯅ @kushal_to]

@cmzw_ Omg

[Kushal ᯅ @kushal_to]

@cagrimbakirci @bcherny Answer is to stop using sub agents.

[Kushal ᯅ @kushal_to]

@RoszykAdam I like this.

[Kushal ᯅ @kushal_to]

@rezoundous Opus high seems to last me about 2 days with my current workflow.

[Kushal ᯅ @kushal_to]

All on opus high.

So far at 14% weekly usage over about 6 hours.

So one 20x should last me about 2 days. 🤞

[Kushal ᯅ @kushal_to]

So far today, creating different sessions with one orchestrator seems to be so much better than one session with sub agents.

[Kushal ᯅ @kushal_to]

@thekitze Even Zelda?

[Kushal ᯅ @kushal_to]

@JustinGorya Yea I’ve noticed this too

[Kushal ᯅ @kushal_to]

@phuctm97 Happy Birthday!

Been great following you this past year!

[Kushal ᯅ @kushal_to]

@ElasticSea Amazing work! Looks so seamless.

[Kushal ᯅ @kushal_to]

Ok this is the way. Not sure why this too me so long to figure out.

[Kushal ᯅ @kushal_to]

It feels like a single Claude is worst at multi tasking vs focusing on one lane at a time.