Superlinear

Podcast on emerging practices for building with AI.

Hosted by @bkase_ and @christinetyip.

[Superlinear @superlinear_fm]

Understand the Loops and Graphs that Frontier Models Like Opus 5, Fable 5, and GPT-5.6 Use

In our new episode:
→ How agent loops become graphs
→ Why focused subagents can perform better
→ Where the extra tokens go
→ When parallelism helps
→ When orchestration becomes expensive overhead
→ How to diagnose a run that gets slow, costly, or stuck

You don’t need to design every graph yourself.

But you should understand the graph your harness is already building.

00:00 Introduction
03:34 How the Agent Loop Works
06:18 Goal Loops and Autoresearch Loops
14:57 From Loops to Graphs
17:06 Sub-Agents: Benefits, Costs, and Context
21:18 The Research Fan-Out Graph
26:27 Multi-Model Agent Graphs
28:37 Asynchronous Implementation and Review
33:40 How to Design Agent Graphs
36:07 When the Harness Builds the Graph for You: Token Spend vs. Performance
38:38 Execution Traces: A Profiler for Agents
41:46 Final Takeaway

[Christine Yip @christinetyip]

https://t.co/T0Eh9biF9C

[Superlinear @superlinear_fm]

For reference, topics we've already covered:

Episode 1:
Point, don't describe: a better way to work with AI agents

https://t.co/KvuH64vc0r

Episode 0:
Four emerging practices for building with AI:
- Leverage shifting left for agents
- Compress intent with expert language
- Match the tool interface to the task
- Use oracles (when you can)

https://t.co/CCgZhFcysJ

[Superlinear @superlinear_fm]

Point, don’t describe, when working with coding agents.

AI tools are moving beyond chat. @karpathy pointed at “audio in, visuals out.” @apike's talk at @aiDotEngineer gestures in the same direction.

Humans think visually. Agents need text context. The workflow unlock is translating between the two.

00:35 Why text is “one-dimensional” and visual interfaces feel higher bandwidth
04:06 The render → annotate → translate loop
05:27 @plannotator for giving feedback on plans and specs
07:00 Claude Desktop, Codex, and Agentation for frontend feedback
08:47 Building spatial feedback mode into a Roblox game
10:38 Why “point instead of describe” may become a core pattern in agent workflows
11:57 What's next: extending this to video, 3D models, simulations, and more

[Superlinear @superlinear_fm]

@eric_khun Both

[Christine Yip @christinetyip]

.@bkase_ and I are always comparing notes on what we’re seeing and trying in AI. We're really glad that people are enjoying the first episodes.

What should we talk about next?

What’s one agentic engineering topic you’d like two AI builders to think and talk through honestly?

[Christine Yip @christinetyip]

Love this.

I always leave our podcast conversations excited to try something new.

Seeing listeners feel the same makes me so happy. It makes all the wrestling with podcast software worth it. 😂

[Hussein Karaki @hussein_builder]

@bkase_ @superlinear_fm Awesome. Im building my first graph now. Taking inspiration from your podcast. Gonna store it in github since I wanna modify the graph over time during Retro - will see how it goes.

[Brandon Kase @bkase_]

Does anyone have thoughts on this?

To what extent are the subagent graphs automatically orchestrated by Opus 5 and Fable and Sol better than ones we try to make ourselves for common tasks?

[Eric - your game multiplayer in 1 prompt @eric_khun]

loop or graph?

[Brandon Kase @bkase_]

Want a specific subagent graph?

Just tell codex about your graph in plain English.

Here I show the graph that I’ve been using to knock out RFCs on my Gamboy game LLM project.

[Superlinear @superlinear_fm]

What do you do when the default orchestration in Opus 5, Fable 5, or GPT-5.6 Sol doesn’t align with your goals?

You design the graph yourself.

We show you how in our latest episode on loops and graphs in agentic engineering. 👇

[Superlinear @superlinear_fm]

Understand the Loops and Graphs that Frontier Models Like Opus 5, Fable 5, and GPT-5.6 Use

In our new episode:
→ How agent loops become graphs
→ Why focused subagents can perform better
→ Where the extra tokens go
→ When parallelism helps
→ When orchestration becomes expensive overhead
→ How to diagnose a run that gets slow, costly, or stuck

You don’t need to design every graph yourself.

But you should understand the graph your harness is already building.

00:00 Introduction
03:34 How the Agent Loop Works
06:18 Goal Loops and Autoresearch Loops
14:57 From Loops to Graphs
17:06 Sub-Agents: Benefits, Costs, and Context
21:18 The Research Fan-Out Graph
26:27 Multi-Model Agent Graphs
28:37 Asynchronous Implementation and Review
33:40 How to Design Agent Graphs
36:07 When the Harness Builds the Graph for You: Token Spend vs. Performance
38:38 Execution Traces: A Profiler for Agents
41:46 Final Takeaway

[Superlinear @superlinear_fm]

@jin_devilfruits Thanks!

[Superlinear @superlinear_fm]

Understand the Loops and Graphs that Frontier Models Like Opus 5, Fable 5, and GPT-5.6 Use

In our new episode:
→ How agent loops become graphs
→ Why focused subagents can perform better
→ Where the extra tokens go
→ When parallelism helps
→ When orchestration becomes expensive overhead
→ How to diagnose a run that gets slow, costly, or stuck

You don’t need to design every graph yourself.

But you should understand the graph your harness is already building.

00:00 Introduction
03:34 How the Agent Loop Works
06:18 Goal Loops and Autoresearch Loops
14:57 From Loops to Graphs
17:06 Sub-Agents: Benefits, Costs, and Context
21:18 The Research Fan-Out Graph
26:27 Multi-Model Agent Graphs
28:37 Asynchronous Implementation and Review
33:40 How to Design Agent Graphs
36:07 When the Harness Builds the Graph for You: Token Spend vs. Performance
38:38 Execution Traces: A Profiler for Agents
41:46 Final Takeaway

[Christine Yip @christinetyip]

https://t.co/T0Eh9biF9C

[Christine Yip @christinetyip]

https://t.co/OqYEflIQEI

[Superlinear @superlinear_fm]

Opus 5, Fable 5, and GPT-5.6 Sol already orchestrate agent graphs for you on their own.

But should you let them?

Or do you still need to design the loops and graphs yourself?

We unpack the trade-offs in our latest episode.

[Superlinear @superlinear_fm]

Understand the Loops and Graphs that Frontier Models Like Opus 5, Fable 5, and GPT-5.6 Use

In our new episode:
→ How agent loops become graphs
→ Why focused subagents can perform better
→ Where the extra tokens go
→ When parallelism helps
→ When orchestration becomes expensive overhead
→ How to diagnose a run that gets slow, costly, or stuck

You don’t need to design every graph yourself.

But you should understand the graph your harness is already building.

00:00 Introduction
03:34 How the Agent Loop Works
06:18 Goal Loops and Autoresearch Loops
14:57 From Loops to Graphs
17:06 Sub-Agents: Benefits, Costs, and Context
21:18 The Research Fan-Out Graph
26:27 Multi-Model Agent Graphs
28:37 Asynchronous Implementation and Review
33:40 How to Design Agent Graphs
36:07 When the Harness Builds the Graph for You: Token Spend vs. Performance
38:38 Execution Traces: A Profiler for Agents
41:46 Final Takeaway

[Brandon Kase @bkase_]

Opus 5, Fable, and Sol are spawning subagents on your behalf. And they’re tuned for the average: the average token budget, the average tolerance for parallelism.

Are you always the average? For all your projects? Maybe sometimes you want to burn tokens to get things done faster. Maybe sometimes you want to save your token plans. Maybe sometimes you need a blend.

Graphs give meaning to otherwise implicit subagent relationships.

You don’t need to design every graph yourself. But you can’t tune what you can’t understand.

In this episode, we build up a mental model for subagent graphs and execution traces from absolute scratch: Starting from agent loops, then loops on loops, then graphs. We tried hard to assume almost nothing about what you already might know.

By the end, you should be able to look at a run that’s slow or expensive or stuck or faulty and actually diagnose where your tokens went and how to make the next run faster and cheaper and smoother and more effective.

[Christine Yip @christinetyip]

If you're using Opus 5, Fable 5, or GPT-5.6 Sol, and care about speed, performance, or token burn,
this episode will help you understand and improve the graphs behind your workflows.

[Brandon Kase @bkase_]

In this episode, we go through foundational material starting from agent loops and discuss loops before we get to reasoning about subagent graphs.

Also, based on feedback, we kept the pauses and disfluencies intentionally to minimize harsh cuts when watching the video.

Let us know what format you prefer and why!

[Superlinear @superlinear_fm]

Subscribe to our newsletter: https://t.co/IhoZTLjfDH
YouTube: https://t.co/DDspcyaIHY
Spotify: https://t.co/ZECdkJTFbV
LinkedIn: https://t.co/iHsGObY0ud

[Christine Yip @christinetyip]

Tried a Specialist Review Graph on my project: 6 subagents, 6 different lenses.

The feedback was noticeably better (much sharper and less generic) right away, because each context stayed focused instead of trying to review everything at once.

Even though Fable and GPT-5.6 Sol often spawn subagents on their own, but giving them a little nudge toward the right graph at the right time for your project can still make a surprisingly big difference.

[Christine Yip @christinetyip]

How to use loops and **graphs** in agent workflows.

[Superlinear @superlinear_fm]

Coming up on Superlinear.

A teaser for our upcoming episode: Loops & Graphs.

[Christine Yip @christinetyip]

https://t.co/T0Eh9biF9C