Discover/This Week in StartupsThis Week in Startups/Behind the Scenes with an early OpenClaw...

Early OpenClaw contributor reveals the token cost tradeoff of AI sub-agents

This Week in StartupsThis Week in Startups · Behind the Scenes with an early OpenClaw contributor! |...

“But, yeah, I I think we should lean towards going towards, like, sub agents just so the main thread is always, like, quick to respond.”

28:19 / 28:57

Is that gonna be automated? This like, shouldn't it know to just

every request on a sub agent? Yeah. Some people don't really want it because, once again, you're using more tokens when you do create a sub agent a sub agent because, I mean, it has to recreate, like, the the memory the prompt and memory and stuff, so it was 20,000 tokens. But, yeah, I I think we should lean towards going towards, like, sub agents just so the main thread is always, like, quick to respond. Yeah. This seems to me like the,

inevitable architecture. We have Mac Studios on order. I understand there's a new Mac Studio coming that's gonna be designed natively

About this clip

An early OpenClaw contributor discusses the architectural decisions around AI sub-agents, explaining why some users resist automation due to increased token costs (20k tokens per sub-agent). The conversation touches on the inevitable shift toward sub-agent architecture for better main thread responsiveness.

Why this clip

Provides inside perspective on real architectural tradeoffs in AI agent systems that other developers are likely facing.

28:19 - 28:5738stechnical insight
ai-agentstoken-costssystem-architecture

Share

LinkedInX

More from this episode

This Week in StartupsThis Week in Startups

tactical advice

AI sub-agents work in background while you chat, then report results

An OpenClaw contributor explains how AI sub-agents can handle complex, time-consuming tasks in the background while users continue chatting normally. The system spawns sub-agents for tasks requiring multiple tool calls, which work independently for 15-20 minutes before reporting results back to the main agent. However, this approach uses more tokens since each sub-agent needs to recreate memory and prompts.

27:55 - 28:3843s
ai-agentsautomation-workflowstoken-optimization

This Week in StartupsThis Week in Startups

market insight

AI companies are growing faster than we've ever seen before

A discussion about Anthropic's valuation reveals that while the numbers seem crazy compared to five years ago, they're actually reasonable when you plot revenue against growth rate. The key insight is that we've simply never witnessed companies growing at this unprecedented pace before.

4:28 - 5:0235s
ai-valuationsrevenue-multiplesstartup-growth

This Week in StartupsThis Week in Startups

prediction

AI will eliminate the lowest-end knowledge work jobs first

The speaker argues that AI's intelligence will displace workers in less intellectual software and tech roles, causing related stocks to drop 25% recently. This raises fundamental questions about what knowledge work means in an AI-dominated future.

37:20 - 37:5434s
ai-displacementknowledge-workmarket-reaction

This Week in StartupsThis Week in Startups

tactical advice

AI models just doubled in speed - but there's still one major bottleneck

An early OpenClaw contributor discusses the dramatic speed improvements in AI models over recent weeks, from 30 to 50-60 tokens per second. The conversation focuses on current technical limitations around speed and threading, specifically the challenge of running multiple parallel tasks with a single AI agent rather than needing multiple instances.

27:03 - 27:4240s
ai-performanceopenclawtechnical-optimization

Similar clips from other shows

The Twenty Minute VC (20VC)The Twenty Minute VC (20VC)

prediction

Building a 'company in a box' with AI agents handling accounting and CRM

Sebastian describes his weekend experiment creating a 'company in a box' system that uses Claude AI agents to manage open source accounting and CRM software. He demonstrates how AI can automate basic business functions like bookkeeping invoices and setting up customer accounts, suggesting a future where companies rely less on traditional SaaS tools.

9:07 - 10:0053s
ai-agentsautomationsaas-disruption

a16z Podcasta16z Podcast · Anish Acharya

market insight

Why Figma is perfectly positioned for the AI era while other SaaS dies

Anish Acharya explains why products like Figma will thrive as AI agents replace execution-focused tools, because they focus on creative thinking rather than just task completion. He argues that 'area under the curve' companies with network effects are better positioned for the AI transition, though they still face fundraising challenges without sufficient momentum.

45:53 - 46:3744s
network-effectsai-agentssaas-evolution

This Week in StartupsThis Week in Startups

prediction

VC plans AI 'replicants' with 200 skills to replace 20 employees

A VC explains their ambitious plan to build AI agents called 'replicants' that combine 100-200 skills to handle the work of 20 people at their firm. The goal isn't mass layoffs but eliminating routine tasks so humans can focus on high-value activities like meeting with founders and LPs.

2:20 - 2:5536s
ai-agentsworkforce-automationvc-operations

The Twenty Minute VC (20VC)The Twenty Minute VC (20VC)

prediction

Why build software when AI agents can connect everything for you?

Sebastian questions whether companies should build integrated software solutions when AI agents will soon move data between products seamlessly. He explores the 'company in a box' model and expresses skepticism about the overcrowded customer support AI space with 14 companies raising $100M+ recently.

12:45 - 13:2541s
ai-agentssaas-integrationcustomer-support

From the blog

Comparison

Clypt vs Opus Clip: Why Generic AI Clippers Fail on VC Podcasts

8 min read

Want clips like this for your podcast?

We find your top 5-8 clips, write the hooks, and deliver ready-to-post content. First 2 episodes are free.

Get 2 Episodes Clipped Free