The Pulse #158: New Ways Of Coding With AI Tools

Gergely Orosz · pdf · 2026-01-09 Reader 🖥️

mixed · mixed — A weekly news-digest newsletter with four largely independent sections (reader feedback/accountability argument, new AI-coding techniques, general industry news, AI-driven job cuts at small software businesses). This extraction focuses on the sections relevant to the team/coding-agents initiative — accountability for AI-generated code, the 'Ralph' autonomous-loop technique, and Claude Code creator Boris Cherny's individual workflow — per the link-roundup guidance to keep digest extractions proportionate to the piece's depth rather than enumerating unrelated industry items (RAM shortages, AWS pricing, Meta/Manus acquisition) exhaustively.

Structure

Key terms

• Ralph
  An autonomous agentic loop technique (coined by Geoff Huntley) that repeatedly re-runs a coding-agent CLI against the same prompt/plan file until a project is done, with no human watching each step — contrasted with the interactive, step-by-step way most developers currently use AI coding CLIs.
• human-in-the-loop (HITL) coding
  (Quoting Matt Pocock) The default mode most developers use today: give the CLI a task, watch it work, intervene when it goes off-track — explicitly contrasted with Ralph's 'long-running, autonomous, unsupervised AFK coding.'
• accountable software engineer
  A person who can be held responsible for software working as expected and, crucially, can determine exactly what went wrong and why if it doesn't — contrasted with a non-technical person who used AI to build something and can only say 'the AI did it' when it breaks.

Propositions

• explicitAs AI agents become more capable, demand for software engineers who can be held accountable for mission-critical software working as expected should increase, not decrease, because more mission-critical software will be built quickly by a wider range of people, and someone has to be answerable for it.conf: high
  supportsResearch: Why You Shouldn't Treat AI Agents Like Employees · p6 — Both argue that because today's AI cannot itself be held accountable, a human must always retain clear ownership; Orosz's CFO/payroll scenario and lawyer anecdote supply a fresh, concrete illustration of exactly the gap this proposition names.
• explicitA non-technical person using AI to modify software they depend on (the CFO/payroll example) cannot actually determine whether the AI's confident-looking output is correct, and has no recourse but to disclaim responsibility after the fact — a failure mode a trained software engineer would avoid because they could isolate and explain the actual defect.conf: high
  supportsPodcast: The AI Joy Gap: Why Some Developers Thrive While Others Struggle · p12 — Both assert that accountability for AI-generated output must stay with a human regardless of who or what generated the code; Orosz extends this from 'developers reviewing AI code' to the sharper case of a non-technical person with no way to review it at all.
• explicitLLMs can produce reasoning on a specialized topic (law, by extension code) that 'looks sound' to someone unversed in the field, while being wrong in ways only a domain expert would catch — because the LLM lacks undocumented, field-specific common knowledge that insiders take for granted.conf: high
• explicitThe 'Ralph' autonomous-loop technique — the same prompt re-run against a coding CLI in a bash loop until a project finishes — only became viable in the weeks before this piece because the underlying AI models improved significantly since November; before that point the same loop would not have produced much working code.conf: high
  supports11 Tips For AI Coding With Ralph Wiggum · p2 — Pocock's piece (which Orosz cites directly as his source) frames Ralph as the terminal stage of a discipline progression from vibe coding to full autonomy; Orosz's observation that the technique only started working in the weeks before publication, due to a model-capability threshold crossed since November, supplies the timing/causal explanation for why that terminal stage suddenly became reachable.
• explicitRunning a coding agent fully autonomously (Ralph/AFK mode) is a qualitatively distinct practice from the default, interactive human-in-the-loop style most developers use — the difference is not degree of automation but whether a human is actively watching and intervening step by step.conf: high
• explicitClaude Code's own creator writes code primarily by running many parallel AI agent sessions himself (5 local, 5-10 remote) rather than by typing — he reports that 100% of the code he personally wrote last month was authored by Claude, prompted and verified by him.conf: high
• explicitA team-wide CLAUDE.md instructional file (e.g. 'always run via this command,' 'always lint before a PR') and giving the agent a way to verify its own work are, in Boris Cherny's account, the most important levers for getting high-quality results — he states that a working verification feedback loop alone can 2-3x result quality.conf: high
  extendsJohn Cutler's Post · p3 — Cutler names 'shared-context solo' (shared repo conventions, but everyone works their own AI session alone) as what most 'team AI' setups actually are; Boris Cherny's reported setup — a team-wide CLAUDE.md plus each developer (himself included) running many of their own parallel, individually-directed Claude sessions — is a concrete, named instance of exactly that pattern from inside the team that builds Claude Code itself.
• explicitRunning multiple parallel AI agents per developer (rather than one agent per developer) is becoming the norm for developers working in 'AI-native' ways, not just an unusual power-user habit.conf: medium
• explicitEven senior, AI-skeptical engineers (David Heinemeier Hansson) distinguish 'pure vibe coding' (still aspirational for professional work) from 'supervised collaboration' (already real and productive) — and resist vendor claims of agents writing 90%+ of code as inconsistent with holding the line on quality and cohesion.conf: high

Arguments

Narratives

• The author needed professional legal advice, got it stamped and in writing specifically so he would be protected and accountable, but while waiting four weeks for the lawyer's answer, ran the same question through Claude, ChatGPT and Perplexity; all three reasoned plausibly but reached the wrong conclusion because they lacked undocumented, insider regulatory knowledge the law firm had.
  → Grounds the accountability argument in a first-person case where confident-sounding AI reasoning was substantively wrong in a way only a domain expert could detect — the evidentiary basis for the CFO/payroll hypothetical that follows.
• After a server ran out of memory and CPU, Tailwind's small, AI-stretched team SSH'd in, installed Claude Code, and used it as an impromptu DevOps engineer to diagnose that load (not a code bug) was the cause.
  → An implicit counter-case to the accountability argument above: a small team without dedicated specialist expertise used an AI agent successfully for exactly the kind of diagnostic work the CFO/payroll example says non-experts can't reliably do with AI — the piece does not reconcile the tension.

Questions addressed

• ✅ Does AI's growing coding capability reduce or increase the market value of a professional software engineer?
  Increases it, specifically for the accountability function — determining why something broke and ensuring it doesn't recur — which the author argues AI and non-technical AI users cannot substitute for.
• ✅ Why did the autonomous 'Ralph' loop technique only become practically viable in the weeks before this piece, rather than whenever the pattern was first conceived?
  Underlying model capability crossed a threshold (model improvements since November) that made unsupervised loops reliable enough to be useful, rather than the technique itself changing.
• ⭕ Is the author's own accountability argument (engineers are needed because non-experts get fooled by confident-sounding AI output) consistent with the Tailwind anecdote in the same piece, where a small team with no dedicated specialist used an AI agent successfully for diagnostic work?

Gaps

• incompleteThe accountability argument and the Tailwind DevOps anecdote sit in the same issue but are never connected: the CFO/payroll hypothetical argues non-experts can't safely rely on AI's confident output, while the Tailwind team (competent generalists, not dedicated DevOps specialists) successfully used Claude Code as an impromptu diagnostic expert. The piece doesn't examine what distinguishes the cases — plausibly, that Tailwind's engineers retained enough general engineering judgment to validate the AI's diagnosis even outside their specialty, which a true non-technical user (the CFO) would lack.
  Read next: Material on what baseline expertise (not necessarily domain-specific) is sufficient to safely supervise AI output outside one's specialty — e.g. the 'expert-generalist' concept.
• incompleteThe piece reports Boris Cherny's individual workflow (parallel Claudes, team-wide CLAUDE.md, verification loops) approvingly as a glimpse of 'AI-native' norms, but never asks whether running 5-10 parallel agent sessions per developer is itself a team-coordination problem — e.g. whether colleagues reviewing Boris's PRs can tell which of his many concurrent Claude sessions produced what, or whether this concentrates review/understanding risk on whoever merges, since no other human was in any one of those loops while it ran.
  Read next: extract-cutler-post-single-player-multiplayer-ai — Cherny's setup (shared team-wide CLAUDE.md infrastructure, but each developer running many of their own solo, parallel agent sessions) is close to a textbook instance of 'shared-context solo,' not multiplayer collaboration.

Related documents

• John Cutler's Post
• Podcast: The AI Joy Gap: Why Some Developers Thrive While Others Struggle
• Research: Why You Shouldn't Treat AI Agents Like Employees
• [FULL WORKSHOP] AI Coding For Real Engineers - Matt Pocock, AI Hero
  Agentic Loop Coding Agents Human-in-the-Loop
• 11 Tips For AI Coding With Ralph Wiggum
  Agentic Loop Human-in-the-Loop
• Software Factories, Light and Dark
  Agentic Loop Human-in-the-Loop
• Have your agent record video demos of its work with shot-scraper video
  Coding Agents Human-in-the-Loop
• Android Studio Quail 2 Redesigns Agent Mode, Streamlines AI-Assisted Coding
  Coding Agents Human-in-the-Loop
• Automated Bug Fixing at Facebook Scale
  Coding Agents Human-in-the-Loop
• Jacob Langvad Nilsson's Post
  Agentic Loop Coding Agents