agentic-coding on tomrochette.com

Harper Reed

tom@tomrochette.com (Tom Rochette) — Thu, 24 Sep 2026 00:00:00 +0000

Harper Reed is a Chicago technologist and entrepreneur (CTO of Obama’s 2012 campaign, former CTO of Threadless, founder of Modest, now CEO of 2389.ai) who writes about shipping software with LLMs at harper.blog. Facts below verified as of 2026-09-24.

His essay “My LLM codegen workflow atm” became the reference spec-first codegen workflow that Hacker News threads cite as best practice, and its boundary is that it is one person’s fast-aging loop, dated on purpose.

What it is
#

A personal blog running about 24 years, roughly 1444 posts at about 1.04 posts per week on average, split into long-form posts, near-daily short notes, and a Now page. The signature piece (2025-02-16) lays out a three-file greenfield workflow: brainstorm a spec with a one-question-at-a-time prompt, have a reasoning model turn spec.md into a prompt_plan.md and todo.md, then execute prompt by prompt with Claude or Aider while tests gate each step, plus a repomix-based context loop for existing codebases. The follow-ups generalize it: “An LLM Codegen Hero’s Journey” maps the nine-step adoption ladder from autocomplete to full agents, and “Basic Claude Code” ports the workflow to Claude Code with TDD as the anti-hallucination device (“The robots LOVE TDD”). Everything ships with the actual prompts inline, and posts carry his “written 98% by a human” disclosure.

Status
#

Active and still iterating in public. As of 2026-09-24 the homepage leads with “Why Don’t My Agents Break Containment?” (2026-09-22), an agent-containment experiment out of his 2389.ai research, with recent long-form posts in January, March, August, and September 2026 and near-daily notes. He is CEO of 2389.ai, where his whole team adopted the Claude Code workflow, and his history page records the arc: CTO of Obama for America from April 2011, Modest sold to PayPal in 2015, 2389 started in 2024. The workflow post’s footprint is verifiable on Hacker News: commenters in the Kiro IDE launch, the Claude 4, and the GitHub Copilot Coding Agent threads point to it as the popular workflow, one calling it “basically best practice today” (as of 2026-09-24). He also translates major posts into Japanese, Spanish, Korean, and Chinese.

Strengths
#

Concreteness: the prompts, the three files, and the exact execution loop are all in the post, and the workflow is tool-agnostic (it outlived his own Aider phase).
Practitioner credibility at scale: many shipped side products, a C interpreter built with Claude Code in about an hour, and a whole team running the process.
Calibrated humility: he dates his own advice (“this is working well NOW, it will probably not work in 2 weeks”) and updated it within weeks when Claude Code shipped.
The follow-ups cover the two gaps most workflow posts leave: how beginners climb to agentic coding, and how a team adopts it through pilots, rotation, and documentation.

Cautions
#

Shelf life is short by his own admission: the original post was Aider-centric, Claude Code landed eight days later, and his current work has moved on to agent containment experiments.
Greenfield and small-team skew: the multiplayer problem (“the bots collide, the merges are horrific”) is named, not solved.
Skeptics exist in the same threads that praise him: one engineer reports never succeeding at building a complete feature or prototype with the technique (visible in the HN results as of 2026-09-24).
The signal is buried in a personal blog: profanity, photography, and a high-volume notes stream mean RSS or targeted visits beat casual scrolling.

Compared to
#

Simon Willison: both practitioner bloggers, but Willison tracks the whole tool-and-security landscape while Reed commits to one opinionated workflow; take breadth from Willison, a starting loop from Reed.
Hamel Husain: Husain covers the evaluation side of LLM systems, Reed covers the generation loop; pair them when your question shifts from “how do I generate” to “how do I know it works”.
Chip Huyen: Huyen’s systems-level ML engineering framing versus Reed’s single-developer daily loop; Huyen for organizations, Reed for the individual.

Bottom line
#

Recommended for developers starting out with agentic codegen who need a concrete, proven workflow to run and mutate, and for leads designing a first team pilot. Not for readers wanting evaluation methodology, organization-level platform strategy, or advice stable enough to survive the next model release.

Changes
#

2026-09-24 - Created.

References
#

https://harper.blog/2025/02/16/my-llm-codegen-workflow-atm/ - the reference workflow post: spec.md, prompt_plan.md, todo.md, and both execution loops
https://harper.blog/ - the blog home: current posts (through 2026-09-22), the notes stream, RSS, and translations
https://harper.blog/about/ - the blogging record: about 24 years, about 1444 posts, about 1.04 posts per week
https://harperreed.com/ - credentials and timeline: Obama 2012 campaign CTO, Threadless, Modest to PayPal, 2389.ai
https://harper.blog/2025/04/17/an-llm-codegen-heros-journey/ - the adoption-ladder follow-up
https://harper.blog/2025/05/08/basic-claude-code/ - the Claude Code port and team adoption
https://harper.blog/2025/04/10/waterfall-in-15-minutes-or-your-money-back/ - the micro-waterfall framing and team pilot advice
https://harper.blog/2026/09/22/break-away/ - the current agent-containment work
https://hn.algolia.com/api/v1/search?query=harper%20reed%20llm%20codegen&hitsPerPage=8 - third-party HN evidence: cited as best practice across Kiro, Claude 4, and Copilot Coding Agent threads, plus one skeptic (as of 2026-09-24)

Owain Lewis

tom@tomrochette.com (Tom Rochette) — Thu, 24 Sep 2026 00:00:00 +0000

Owain Lewis is a UK-based AI engineer and engineering director whose YouTube channel walks through the agent infrastructure he builds and open sources. Facts below verified as of 2026-09-24.

He is the category’s build-and-ship voice: every video is a working system (a software factory, a coding agent, a personal agent gateway) with the repo linked in the description, which makes him the member to clone from, and the boundary is that his scope is one-person and consultancy scale.

What it is
#

A YouTube channel (18.7K subscribers as of 2026-09-24) run by Owain Lewis, who describes himself as a principal AI engineer and engineering director with 20 years in software, now running his own AI consultancy. Coverage is practical agentic engineering: agent loops, nested subagents, software factories, coding agents built from scratch, and Claude Code and Codex workflows. Each video pairs with open source he maintains: Machinist (factory infrastructure, 455 stars), blueprint (agent skills, 403 stars), Push (a coding-agent gateway, 196 stars), Neo (a minimal multi-agent harness, 144 stars), and a youtube-tutorials companion repo. Uploads run about weekly, and descriptions read like engineering writeups that state token costs, failure modes, and when he would not use his own build.

Status
#

Active and steady at modest scale as of 2026-09-24. The channel was created 2014-07-20 and shows 18.7K subscribers; the RSS feed shows 15 uploads between 2026-05-15 and 2026-09-21, roughly one per week. The latest upload (2026-09-21, “Inside OpenAI’s Agentic Software Factory”) had 12,960 views and 180 ratings within three days. Reach concentrates in the factory videos: “I Built an Agentic Software Factory” (2026-07-25) is his biggest recent at 68,787 views, while most uploads land between 3,000 and 20,000. Trajectory is upward without a breakout, and his roughly 1.2K combined GitHub stars across the pinned repos carry more durable weight than the sub count suggests.

Strengths
#

Every claim comes with a repo: the factories, agents, and gateways he demonstrates are runnable open source, not concept demos.
Descriptions carry the numbers creators usually cut: around $2 of API spend in two days on the DeepSeek harness test, and explicit statements of when a build is the wrong tool.
His testing surfaces findings the vendors announce differently: he nested Claude Code subagents roughly 30 levels deep after the announced cap was five, and showed Codex’s max_depth guardrail as the fix.
The weekly cadence follows one coherent thread (factory thinking applied to a single engineer’s workflow), so older videos stay usable instead of expiring.

Cautions
#

Every description funnels to his aiengineer.co email capture, pitched with “$1M+ in software deals” credits, so treat the free-skills funnel as the business model it is.
The scope is one-person consultancy scale: his factory runs his backlog, not a regulated enterprise codebase, so transplant the patterns rather than the numbers.
No critical community filter exists: Hacker News has zero threads on his channel content (searched 2026-09-24), so his techniques have not been stress-tested in public argument.
View counts are modest, so the community surface for catching errors in his techniques is thin.

Compared to
#

IndyDevDan: both build software factories in public, but Dan publishes weekly concept frameworks with louder packaging while Owain publishes quieter end-to-end builds you clone directly.
AI Jason: both walk through complete working workflows, but Jason’s builds target everyday app use cases while Owain’s target the engineering loop itself.
Caleb Writes Code: use Caleb for same-week release explainers and Owain when you want to see a system built and open sourced around one.

Bottom line
#

Recommended for engineers who learn by cloning: his value is the linked repos and the stated tradeoffs, not the watching. Not for someone who wants model-release news, enterprise-scale case studies, or a large community already testing the techniques.

Changes
#

2026-09-24 - Created.

References
#

https://www.youtube.com/@owainlewis - the channel: identity and 18.7K subscribers as of 2026-09-24
https://www.youtube.com/feeds/videos.xml?channel_id=UC08YjBHjjPFu8T1KbHGomnQ - the RSS feed grounding cadence, titles, dates, view counts, and the 2014-07-20 channel creation
https://github.com/owainlewis - the GitHub profile linking this channel, with bio, location, and pinned repo star counts
https://owainlewis.com - the personal site: positioning, newsletter, and community links
https://aiengineer.co/start - the free starter pack and email funnel behind every video description
https://hn.algolia.com/api/v1/search?query=%22owain%20lewis%22&hitsPerPage=8 - zero Hacker News hits for the name, the missing critical coverage
https://hn.algolia.com/api/v1/search?query=owainlewis&hitsPerPage=5 - 82 hits with the oldest visible from 2014 to 2015 (Forth links, awesome lists, early startups), grounding his long engineering history and the absence of current channel discussion

Ray Amjad

tom@tomrochette.com (Tom Rochette) — Thu, 24 Sep 2026 00:00:00 +0000

Ray Amjad is an engineer-educator (ex-YC technical founder, Cambridge physics) whose YouTube channel turns each Claude Code and Codex release into a worked explanation with a verification-first argument attached. Facts below verified as of 2026-09-24.

He is the fastest way to understand a Claude Code feature drop in depth, and the only channel in this batch that states it has never accepted a sponsor, with the boundary that his videos are also the top of a cohort funnel and his own products are his demos.

What it is
#

A YouTube channel (49.8K subscribers as of 2026-09-24) run by Ray Amjad, who builds agent-first software under his holding company 21 Dreams (AgentStack, Impello AI, and the open-source HyperWhisper). Coverage is Claude Code centric: hooks, forked and nested subagents, output styles, dynamic workflows, and the monitor tool, each explained with a working demo within days of release, plus concept videos on loop engineering and verification. Every description states that he has never accepted a sponsor and that his own products keep the channel running. The funnel is agenticcoding.school: a two-week Agent-Era Engineer cohort (237 lessons, next run 2026-09-28 to 2026-10-09), evergreen Claude Code and Codex classes, company training, and consulting.

Status
#

Active and growing as of 2026-09-24. The channel was created 2019-08-09 and shows 49.8K subscribers; the RSS feed shows 15 uploads between 2026-04-09 and 2026-09-18, a steady two to three per month with bursts around Anthropic releases. Reach is concentrated and rising: “Jev + Claude Code = The Cheapest Agentic Coding Loop Yet” (2026-09-18) passed 150,259 views within a week, the Opus 5 output-styles video passed 102,888, and typical uploads run 10,000 to 46,000. His cohort page claims 4,000+ engineers taught and 5,000+ hours in Claude Code and Codex, self-reported figures I found nowhere else.

Strengths
#

Depth per release is his edge: he reads the change, surfaces the undocumented details (for example the function-hooks environment variable when hooks shipped disabled by default), and demos edge cases in one video.
The verification thesis is a real argument rather than a mood: generation is solved so the bottleneck moved, prefer a verifier over an instruction, and merge gates that scale with blast radius.
He watches adjacent experiments (cmux peer sessions, Codex managed threads, Anthropic observer agents) and connects them into patterns before they are mainstream.
With no sponsor reads, there is no advertiser incentive to stretch a claim, though the cautions below show where his incentive actually sits.

Cautions
#

Every video ends in the cohort pitch: the channel is a content layer for agenticcoding.school, so the free material is the introduction, not the product.
His products are his demos: videos touching AgentStack, Impello, or HyperWhisper blur the line between coverage and advertisement for his own stack.
The credentials (4,000+ taught, 5,000+ hours, “most in-depth Claude Code content on YouTube”) are self-reported and unverified by any third party I could find.
The third-party footprint is thin: zero Hacker News threads mention him as of 2026-09-24, so nothing he claims has yet been publicly criticized or corrected, which cuts both ways.

Compared to
#

IndyDevDan: both sell cohort courses and coin vocabulary, but Ray is Claude Code release first and verification first while Dan is harness first, model-stacking first, and weekly.
Caleb Writes Code: both cover Claude Code updates fast, but Caleb compresses to short illustrated explainers while Ray goes deep on one feature with a setup you can copy.
Boris Cherny: Boris is the Claude Code creator speaking from inside the tool; Ray is the most dedicated outside analyst of the same surface.

Bottom line
#

Recommended for engineers who want every Claude Code release explained with depth and a verifier-first mental model, and who can discount the course funnel. Not for someone who wants vendor-agnostic coverage, or who minds that the demos showcase his own products.

Changes
#

2026-09-24 - Created.

References
#

https://www.youtube.com/@RAmjad - the channel: identity and 49.8K subscribers as of 2026-09-24
https://www.youtube.com/feeds/videos.xml?channel_id=UCLA7cJBnqr0nLF2bQBD9uUg - the RSS feed grounding cadence, titles, dates, view counts, the no-sponsor statement, and the 2019-08-09 channel creation
https://www.rayamjad.com/ - the personal site: 21 Dreams studio, consulting, and company registration
https://github.com/ray-amjad - the GitHub profile (Tokyo) whose pinned repos match the video-description links, closing the identity loop
https://www.agenticcoding.school/cohort - the cohort page: positioning claims, syllabus, products, and the commercial model
https://hn.algolia.com/api/v1/search?query=%22ray%20amjad%22&hitsPerPage=8 - zero Hacker News hits for the name, the thin third-party footprint

agentic-coding on tomrochette.com

Harper Reed

What it is
#

Status
#

Strengths
#

Cautions
#

Compared to
#

Bottom line
#

Top 5 recommended reading
#

Changes
#

See also
#

References
#

Owain Lewis

What it is
#

Status
#

Strengths
#

Cautions
#

Compared to
#

Bottom line
#

Top 5 recommended reading
#

Changes
#

See also
#

References
#

Ray Amjad

What it is
#

Status
#

Strengths
#

Cautions
#

Compared to
#

Bottom line
#

Top 5 recommended reading
#

Changes
#

See also
#

References
#

agentic-coding on tomrochette.com

Harper Reed

What it is #

Status #

Strengths #

Cautions #

Compared to #

Bottom line #

Top 5 recommended reading #

Changes #

See also #

References #

Owain Lewis

What it is #

Status #

Strengths #

Cautions #

Compared to #

Bottom line #

Top 5 recommended reading #

Changes #

See also #

References #

Ray Amjad

What it is #

Status #

Strengths #

Cautions #

Compared to #

Bottom line #

Top 5 recommended reading #

Changes #

See also #

References #

What it is
#

Status
#

Strengths
#

Cautions
#

Compared to
#

Bottom line
#

Top 5 recommended reading
#

Changes
#

See also
#

References
#

What it is
#

Status
#

Strengths
#

Cautions
#

Compared to
#

Bottom line
#

Top 5 recommended reading
#

Changes
#

See also
#

References
#

What it is
#

Status
#

Strengths
#

Cautions
#

Compared to
#

Bottom line
#

Top 5 recommended reading
#

Changes
#

See also
#

References
#