Thread-Based Engineering: How to Ship Like Boris Cherny

Key Takeaways

How do you know you're improving as an agentic engineer? IndyDevDan proposes thread-based engineering—a framework for thinking about your work with AI agents as discrete "threads" of work where you show up at the prompt and the review.

The Base Thread

A thread is a unit of engineering work over time driven by you and your agents. Every thread has two mandatory nodes:

You prompt or plan at the beginning
Your agent works through tool calls
You review or validate at the end

Tool calls roughly equal impact. Pre-2023, engineers were the tool calls. Now agents handle the middle while you show up at the beginning and end.

The Six Thread Types

Thread	Name	Description
Base	Single thread	One prompt → agent work → one review
P	Parallel	Multiple threads running simultaneously (Boris runs 5-10 Claude Codes in parallel)
C	Chained	Multi-phase work with intentional checkpoints for high-risk production work
F	Fusion	Same prompt to multiple agents, then aggregate the best results
B	Big/Nested	Meta-threads where agents prompt other agents (sub-agents, orchestrators)
L	Long	High-autonomy, extended duration work—hundreds of tool calls over hours

There's also a hidden seventh: the Z-thread (zero touch)—maximum trust where you skip the review step entirely.

The Framework

flowchart LR
    subgraph Base["Base Thread"]
        P1[Prompt] --> TC[Tool Calls] --> R1[Review]
    end

    subgraph Scale["Scaling Dimensions"]
        More[More Threads]
        Longer[Longer Threads]
        Thicker[Thicker Threads]
        Fewer[Fewer Checkpoints]
    end

    Base --> Scale

Four Ways to Improve

Run more threads — Spin up multiple terminals, work trees, or cloud instances
Run longer threads — Better prompting and context engineering enable higher autonomy
Run thicker threads — Nest threads inside threads (sub-agents, orchestrators)
Reduce checkpoints — Build systems you trust so you can skip human-in-the-loop reviews

Boris Cherny's Setup (Referenced)

5 Claude Codes in terminal (tabs 1-5)
5-10 additional Claude Codes in the web interface for background work
Uses stop hooks for very long-running tasks
System notifications when Claude needs input
Always uses Opus 4.5

Notable Quotes

"Tool calls roughly equal impact assuming you're prompting something useful. Pre-2023, you and I were the tool calls."

"The engineer kicking off five agents in five separate terminals is getting more done than the one running a single agent."

"If you don't measure it, you will not be able to improve it."

Connections

ralph-wiggum-technique-guide — The Ralph Wiggum pattern mentioned throughout is a specific implementation of thread-based engineering (loop over agent for deterministic work)
boris-cherny-on-what-grew-his-career-and-building-at-anthropic — Boris's actual interview where he discusses his Claude Code philosophy and the generalist engineering approach
building-effective-agents — Anthropic's guide covers similar patterns: parallelization maps to P-threads, orchestrator-workers maps to B-threads

Connections (23)