Superlinear

How to Build AI Agents Like Dots, Instinct, Muse & Grok Bot

Wed, 07 Oct 2026 10:00:00 -0400

How do you build an agent like Dots, Muse, Instinct, or Grok Bot? Start high up the stack, and move down only when you need more control.

YC F26-startup Agent 37 founder, Vishnu Krishnaprasad, explains when to use Hermes, OpenClaw, or Pi, why each user needs a separate sandbox, and how to turn a proven manual workflow into skills. He also demos a Dots-like agent built on Hermes.

Superlinear is a podcast about emerging practices for building with AI.

Hosted by Brandon Kase and Christine Yip.

Homepage: ⁠⁠https://superlinear.fm
⁠⁠Twitter: ⁠⁠https://x.com/superlinear_fm
⁠⁠LinkedIn: ⁠⁠https://www.linkedin.com/company/superlinearfm⁠⁠
Brandon Kase: ⁠⁠https://x.com/bkase_⁠⁠
Christine Yip: ⁠⁠https://x.com/christinetyip⁠

Music licensed through Soundstripe. Code: O2UT6CCEQNH4KT0D

How to Build and Verify Agent-Written Code with Bend 2

Fri, 02 Oct 2026 06:00:00 -0400

Bend2 puts code and proofs in the same language, so coding agents can write both. Tests check examples. A checked proof can establish that a specific rule holds for every allowed input.

This is formal verification, an additional verification layer alongside types and tests. You still need to review whether the rule captures the behavior you actually want.

In this episode, we show how to choose rules worth proving, get the agent to finish the proofs, and decide where the extra effort pays off. We uncover false proposed rules and integrate a Bend core with checked editing rules into an existing TypeScript app. You can start with one important part of your own project.

Jev and Decision Models: A New Layer for AI Agents

Thu, 24 Sep 2026 08:00:00 -0400

Your coding agent does not need a frontier LLM call for every judgment. Deciding which tests to run, which skill to load, or whether a prompt needs more context can be a bounded question.

Jev and other decision models return typed choices, scores, or yes/no probabilities quickly enough to sit inside repeated workflows. A valid answer can still be wrong, so placement and fallback rules matter.

We show a live prompt checker and a small, six-commit test-selection experiment, then explore agent routing, local Laya, and how to choose between code, a decision model, a generative model, or a human for each step.

The Fall 2026 Workflow for Starting Projects with Coding Agents

Fri, 18 Sep 2026 07:00:00 -0400

GPT-6 Astra and Fable 5.1 are changing how we start new software projects:

Less harness engineering, less process, less prescribing. More deliberate steering on the few decisions that compound through the whole project.

We walk through the Codex conversation of an actual project that we built: where we applied engineering judgment, what we left to the model, and how we planned verification so the agent could detect mistakes and keep building autonomously.

From the first ramble to the stack and architecture decisions, this is our Fall 2026 workflow for starting a project with coding agents.

Resources mentioned

Matt Pocock — Grill with Docs The grilling skill Brandon uses to have the coding agent interview him, sharpen decisions, and build shared context before implementation. Grill with Docs

Kit Langton Effect educator and practitioner whose work Brandon references for Effect patterns and tooling. https://kitlangton.com/ https://x.com/kitlangton

Effect The TypeScript framework Brandon uses for typed errors, dependencies, concurrency, and application logic. Effect documentation

Effect Solutions Practical patterns and reference material for building applications with Effect. https://www.effect.solutions/

Effect Institute Interactive resources for learning Effect. https://effect.institute/

Alchemy TypeScript infrastructure-as-code built around Effect, which Brandon uses to make infrastructure part of the codebase the agent can inspect and modify. Alchemy

Lexi Lambda — “Parse, don’t validate” The type-driven design article behind the “make impossible states unrepresentable” principle discussed in the episode. Parse, don’t validate

Rich Sutton — “The Bitter Lesson” The essay behind the “Bitter Lesson” idea we refer to when discussing why more capable models can make some bespoke scaffolding less useful. The Bitter Lesson / Rich Sutton’s publications

New Model Dropped? 4 Ways to De-slop Your Project

Fri, 11 Sep 2026 08:00:00 -0400

Your project still carries decisions made before the latest model arrived: tests for removed features, slow checks, outdated code, and instructions that keep accumulating.

A new model is a useful prompt to revisit those decisions. Give your agent a concrete problem to investigate, then review what its proposed changes would improve or put at risk.

In this episode, we cover four maintenance tasks through our own projects: pruning low-value code and tests, turning recurring mistakes into lint rules, reassessing implementations, and updating agent instructions with official model guidance. Examples include an editor cleanup that removed about 6,000 net lines across 131 files, a custom Effect lint rule, and a solver investigation that cut round-trip time from 20 seconds to 1.5 seconds.

Give Coding Agents Eyes Beyond the Chat: Screen Recordings, Session History, and Devices

Thu, 03 Sep 2026 05:00:00 -0400

Your coding agent can inspect its own chat, but it cannot see you switching between browser research, documents, local files, and devices.

Screen recordings can turn that invisible work into model context effectively with video native models like Gemini.

In this episode, we show how Screenpipe exposed Christine acting as the context bridge around Codex, and Brandon's workflow for recording focused work sessions, compressing them to 720p at one frame per second, splitting them into 15-minute clips, and asking Gemini for a detailed play-by-play. We also extend the same idea to real phone test rigs, cameras, friction logs, and skills generated from recorded tasks.

Make Your Coding Agents Expert Designers

Fri, 28 Aug 2026 07:00:00 -0400

AI can generate a website in seconds. So why does so much AI-generated design still look generic?

The problem is not the model. It’s the workflow.

In this episode, we break down three real things we built: the Superlinear website, a custom minimap for our podcast editor, and a 3D game character. We show how breaking design into intermediate steps can give you better results and more places to steer.

Context Hygiene for Coding Agents: Murder, Lobotomy, or Diary

Tue, 18 Aug 2026 10:00:00 -0400

A million-token context window doesn’t mean you should fill it.

Every piece of information you keep in a coding agent’s active context has a cost. Not just in tokens and dollars, but in the model’s attention.

In this episode, we explore context hygiene for coding agents: how to decide what deserves to stay in active context, what to move elsewhere, and when to throw context away entirely.

Beyond Root-Level Harness Engineering: A Different Environment for Every Task

Wed, 12 Aug 2026 10:00:00 -0400

Harness engineering usually happens at the root level of a project.

However, it should go much deeper.
Frontend work, debugging, performance optimization, reviews, and other hard tasks can each have their own purpose-built environment, and even sub-agents inside those tasks can have different harnesses.

4 Emerging Practices in Agentic Engineering

Tue, 28 Jul 2026 12:00:00 -0400

Coding agents are becoming more capable, but getting better results increasingly depends on the systems, interfaces, and feedback loops we build around them.

In this episode of the Superlinear podcast, we unpack four emerging practices in agentic engineering:

→ Shifting left in your agent harness
→ Using precise expert language to communicate intent
→ Choosing deliberately between MCP, CLI tools, Bash, and Code Mode
→ Using test oracles to help agents explore, verify, and reproduce complex behavior

We discuss ideas from AI Native DevCon London, and explore how better guides and sensors can improve a harness, why expert vocabulary helps you access more of an LLM’s capabilities, when code is a better interface than individual tool calls, and how working examples can communicate intent more precisely than a written specification.

The broader takeaway: building effectively with AI is not only about choosing a better model or writing a better prompt. It is also about designing a better environment for the agent to work within.

Understanding Loops and Graphs in Modern Coding Agents

Fri, 24 Jul 2026 14:00:00 -0400

When you ask Fable or GPT-5.6 Sol to complete a task, it feels like you’re talking to one agent.

But behind that single conversation, the harness may be spawning multiple sub-agents, giving each one a fresh context window, running them in parallel, and passing summaries between them.

You gave one prompt. It built a graph.

In this episode of Superlinear, Christine Yip and Brandon Kase unpack how modern agent harnesses grow from simple agent loops into hidden graphs of sub-agents, and why those graphs may be one reason your token limit disappears so quickly.

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

Point, Don't Describe: A Better Way to Work With AI Agents

Fri, 24 Jul 2026 09:00:00 -0400

Most AI workflows still happen in one-dimensional chat. But the best tools are moving toward higher-dimensional interactions: visual, spatial feedback loops where humans can point instead of describe.

00:00 Intro
00:35 Why text is “one-dimensional” and visual interfaces feel higher bandwidth
04:06 The render → annotate → translate loop
05:27 Planotator for giving feedback on plans and specs
07:00 Agentation, Claude, and Codex 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