These sources argue about how programming changes when agents write much of the code. Most are essays and discussion threads. This page lists 32 sources, oldest first where dated.
Disciplined Methodology: Constraints, checks and skepticism for building software with LLMs
The README introduces the method, six chapters and a related format called PAG. The chapters are separate files that this page only summarizes. It ends with an FAQ in which the author describes how the method developed.
Tools: Code Is All You Need
A personal blog post from a software developer who has tried MCP in agentic coding. Its main evidence is one account of converting a blog through generated scripts. The post gives no benchmarks, only the author's own experience. It also suggests a quick GitHub MCP versus gh CLI comparison for readers to run.
Should we revisit Extreme Programming in the age of AI?
The post argues from a few cited statistics and from XP's stated values. It links its view to the firm's Product Operating Model. It reports no tests of XP with AI tools.
Software Development in the Time of Strange New Angels
Dave Griffith opens the newsletter with a claim that coding cost shaped every software practice. The post tells one story of Claude writing a property-based test suite in fifteen minutes. It then predicts that knowing what to build becomes the bottleneck.
Code like a surgeon
The post offers a metaphor and a list of tasks the author hands to agents. It rests on the author's own working habits. The author discloses working at Notion, a maker of AI-related products.
How to Fix Any Bug
The post teaches a method for fixing bugs. The steps are find a repro, narrow it, then remove code. It tells the method through one real bug and one Claude session. The evidence is a single case plus the author's past experience.
Ask HN: How can I get better at using AI for programming?
An Ask HN question with a long run of practical advice from commenters. Early replies cover specs, small tasks, tests, voice input and which tools to use. The evidence is personal workflow, with no measurement. Only the opening part of the thread was read.
AI vs Human Code Generation Report
The post is a summary of a longer report by an AI code review company. It gives rates of issues per pull request and a ranked list of ten findings. It ends with advice and a pitch for the company's product.
Your job is to deliver code you have proven to work
The post sets a standard for contributions: manual proof plus an automated test. It extends the standard to coding agents. It rests on the author's own practice and offers no data.
Build Software. Build Users
The post argues that software quality comes from knowing users, not from test volume. It proposes a folder of user profiles and happy-path flow files. Agents would then play those users against the software. It offers no trial of the idea.
The Future of Software Development is Software Developers
A Hacker News thread on a post arguing that AI will not end programming jobs. The first replies cover model costs, whether LLMs understand, and worry about jobs. Evidence is argument and anecdote. Only the opening part of the thread was read.
Immutable Infrastructure, Immutable Code
Fowler carries the immutable-infrastructure rule over to application code. The post claims AI makes regeneration cheap, so in-place edits become the risky choice. It offers analogy and a short personal anecdote. It gives no data and no regeneration trial.
The Next Two Years of Software Engineering
Osmani frames five questions: juniors, skills, roles, specialists and education. Each gets a pessimistic and an optimistic scenario. The post says these are lenses for preparing, not predictions. Each section ends with advice for junior and senior developers.
--dangerously-skip-reading-code
The post follows an earlier one by the same author. It explores a hypothetical in which leadership accepts the risks of unread code. It offers argument and a sketch of a process, with no tests or data.
We Upgraded to a Frontier Model and Our Costs Went Down
The post describes the agent design behind a CI-failure product. It gives some usage figures and one worked incident. The figures are the company's own and are not independently checked.
Checking an LLM's work is a systemic, not an individual, problem.
The essay is aimed at product and design practice more than at programming. It leans on quotes and links to other writers. It offers argument and no measurements.
slow ai: ai that matches a human's pace
Jenny Huang argues that fast, fluent AI answers encourage acceptance without questioning. The essay offers four design ideas and two prototypes. It reports that study participants felt more ownership with a decision-tree interface.
Lessons for Agentic Coding: What should we do when code is cheap?
A Hacker News thread about a short list of tips for coding with agents. Commenters argue over cost, juniors, review and whether code is really cheap. The evidence is mostly personal anecdote, with a few uncited figures. Opinions range from enthusiastic to hostile.
Vibe coding and agentic engineering are getting closer than I'd like
A very large Hacker News thread on a short blog post from a well-known programmer. Commenters argue over code quality, rewrites, market forces and who reviews code. The evidence is anecdote and prediction. Only the opening part of the thread was read.
Using AI to write better code more slowly
A short personal blog post with a comment thread that includes replies from the author. It describes one reviewer skill that runs three AI reviewers on a pull request. The evidence is the author's own experience, with no measurements. Commenters add their own workflows and some disagreement.
We should be more tired than the model
A Hacker News thread about a short personal essay on agentic coding and memory. Commenters share workflows, argue about skill loss and debate IDE tools versus agents. Evidence is mostly first-hand experience.
The Speed of Prototyping in the Age of AI
Daryl Cecile reflects on a year of working with coding agents. The note gives one rough self-measured figure and a list of new repos. It also names a skill loss the author worries about. It links to other engineers' accounts.
AI demands more engineering discipline. Not less
Majors links the shift to immutable infrastructure with AI code generation. The post leans on Chad Fowler's Phoenix Architecture writing. It argues that validation should move to tests, evals and observability in production. Its support is personal history, quotation and argument.
The primary purpose of code review is to find code that will be hard to maintain
A Hacker News thread about a two-paragraph Mastodon post. Many commenters disagree that review has one main purpose. Many list other purposes, and a long side argument covers the meaning of one phrase. The evidence is workplace experience and argument.
The Shape of Things to Come, Part 1: The Continuous Thunderdome
The essay is a first-person tour of Wheelhouse, a closed-source agent harness. The author built it for an old online game. Its evidence is the author's own usage figures and anecdotes. It makes predictions about code review, CI/CD and automated request handling. It also previews a Part 2 on model welfare.
My company now has like 6 developers all developing like 2-3 substantial features at a time
This is a single short post. It says a six-developer company builds several features at once. Its developers complain about review volume. It has one reply. It offers no data and does not say what produces the code.
Trying the Software Factory pattern.
Will Larson describes the software factory pattern. It loops on a broad goal with a harness. The post lists the steps of a first agent skill. The skill works from a Linear project. It reports the skill works well enough, without data.
Evolving programming languages in the AI era
José Valim splits the post into reflections and tooling ideas. The reflections ask open questions about communities, syntax and compilers. The tooling half makes specific proposals, such as program databases and runtime observability. The proposals draw on experience building Tidewave and on Elixir.
Coding is NOT solved
Ewerlöf argues against the claim that coding is solved. The post is a series of short sections, with screenshots, memes and later updates. Its evidence is the author's experience, quoted posts and reports of product bugs. The tone is polemical and names companies and people.
Control the ideas, not the code
antirez argues that reading AI-written code line by line is often pointless. The post is partly a reply to a question about reviewing Redis code. It draws on the author's work on DwarfStar and Redis. It offers experience, not measurements.
A new era for software testing
antirez describes asking an LLM agent to run release QA from a markdown file. The post gives two projects as examples, DwarfStar and Redis Arrays. It offers the author's experience and expectations. It reports no results or bugs found.
The Specificity Gradient, and Other Things
Dorian Taylor argues that code holds too much detail to be where business decisions live. The essay sets out six layers, from business ecosystem to code. Finer layers change faster than coarser ones. It ends with a vision of a structured intranet.
Glossary
- 4GL
- A fourth-generation language, a high-level tool once billed as replacing programmers.
- Accountable
- Answerable in law or at work for the results, unlike merely responsible.
- Agentic AI
- AI that carries out multi-step tasks, such as writing and running code.
- Agentic code generation
- Having an AI agent write code on a person's behalf.
- Agentic coding
- Using an AI agent that edits code and runs commands on its own.
- Agentic engineering
- The term the post uses for professional software work done with coding agents.
- Amdahl's law
- Speeding up one step of a process gains little if other steps stay slow.
- AST
- An abstract syntax tree, the structure a parser builds from source code.
- Autonomy slider
- Andrej Karpathy's idea that AI tools range from tightly supervised to fully independent.
- Ball of mud
- A codebase with no clear structure.
- BDD
- Behaviour-driven development, writing tests as plain descriptions of expected behaviour.
- Beads
- An issue tracker the author uses as a shared task graph for agents.
- Bisection
- Splitting a failing batch in half repeatedly to find the bad change.
- Blast radius
- How much of a system a single failure can damage.
- Bureau of Labor Statistics
- A United States government agency that publishes employment data.
- Bytecode
- Low-level instructions that a language runtime executes, rarely read by people.
- CHAOS report
- A long-running Standish Group survey of IT project success rates.
- CI
- Continuous integration: automated builds and tests that run when code changes.
- CI/CD
- Automated systems that test and release code changes.
- ClickHouse
- A database built for fast queries over large volumes of data.
- CLIRunner
- A tool in the Python Click library for testing command-line programs.
- Code review
- Having another person read a proposed code change before it is merged.
- Codex
- OpenAI's coding agent.
- Coding agent
- An AI tool that can run code and edit files to complete programming tasks.
- Configuration drift
- Slow, untracked divergence of a running system from its intended setup.
- Context
- The text a language model can see and work with at one time.
- Context window
- The amount of text a language model can consider at once.
- Cursor
- A code editor with built-in AI agents.
- Cursor Bugbot
- An automated bug-finding reviewer made by Cursor.
- Design critique
- A structured review where a team questions why a design decision was made.
- DESIGN.md
- A markdown file that describes how a program is designed.
- Deskilling
- The loss of core abilities when a tool does the work.
- Distillation
- Training a small model to imitate a larger one.
- Dunning-Kruger effect
- The tendency of people who know less to be more confident.
- E2E test
- A test that exercises a whole application from the user's side.
- Elixir
- A programming language that runs on the Erlang virtual machine.
- End-to-end test
- A test that checks a whole product's behaviour rather than its inner parts.
- Eval
- A test that scores how well an AI system performs a task.
- Evaluation
- A check that tells whether a new version behaves correctly.
- Extreme Programming
- A 1990s software method built on pair programming, small releases and constant testing.
- False positive
- A reported bug that is not actually a bug.
- Gh CLI
- The GitHub command-line tool.
- Git worktrees
- A Git feature that checks out several branches in separate folders at once.
- Hacker News
- A technology news and discussion site run by Y Combinator.
- Halting problem
- A proven limit showing no general method can tell whether every program will stop.
- Happy path
- The steps a user takes when everything goes right.
- Harness
- Code that wraps a language model in a loop with tools and checks.
- Human-in-the-loop
- A step in a process where a person must review or approve.
- Hyper-scale LLM
- A very large language model that costs a great deal to train and run.
- Immutable infrastructure
- Servers that are replaced with fresh copies, never changed while running.
- Inference
- Running a trained language model to get its output.
- Inference engine
- Software that runs a trained language model to produce output.
- Integration test
- A test of several parts of a system working together.
- Jevons Paradox
- The idea that making a resource cheaper to use can raise total use of it.
- Knowledge graph
- A web of facts stored as linked items and relations.
- Linear
- A project and issue tracking tool.
- Linting
- Automated scanning of code for style problems and likely mistakes.
- LLM as a judge
- Using a language model to assess the output of a process.
- LSP
- The Language Server Protocol, a way for editors to offer code tools such as rename.
- Mastodon
- A decentralised social network.
- Max account
- A paid Claude subscription tier with high usage limits.
- MCP
- Model Context Protocol, a standard for connecting AI tools to outside data.
- Merge queue
- A line of proposed code changes waiting to be tested and merged.
- Model checking
- Checking a system against a model of its behavior to find violations.
- Multiverse analysis
- A method that runs every reasonable set of analysis choices and reports each result.
- Non-functional requirements
- Qualities such as reliability, security and scalability, apart from what the software does.
- Normalization of deviance
- Slowly accepting risky shortcuts because earlier ones caused no harm.
- Normative truth
- A statement of what should be the case, as opposed to what is.
- Observability
- The ability to understand a running system from the data it emits.
- Open-weight model
- An AI model whose trained parameters are published for anyone to run.
- Orchestrator
- A developer who designs a system and assigns tasks to AI agents and services.
- Pace layers
- A model where parts of a system change at different speeds.
- PAG
- Pattern Abstract Grammar, the author's structured format for writing LLM instructions.
- Pair programming
- Two developers working together at one workstation.
- Plan file
- A written plan an AI agent reads before it starts coding.
- Playwright
- A tool that controls a web browser from code.
- Playwright MCP
- A tool that lets an AI agent drive a real web browser.
- Policy-as-code
- Rules such as style checks written as automated checks.
- PR
- A pull request, a proposed change to a codebase.
- Program database
- A store of facts about code, such as definitions and calls, that can be queried.
- Property-based testing
- Testing code against rules that must always hold, using many random inputs.
- Pull request
- A proposed code change submitted to a project for review.
- QA
- Quality assurance, the checking of software before it is released.
- Quality gate
- The full set of checks a change must pass before it is accepted.
- Rate ratio
- A comparison of how often something happens in one group versus another.
- Refactoring
- Restructuring code without changing what it does.
- Regression
- A change that makes something worse that used to work.
- Repro
- Short for reproduction: a reliable set of steps that shows a bug happening.
- Review
- Reading another person's code changes to find problems before they are accepted.
- RFC
- A written proposal document, here kept in Notion.
- SAST
- Static application security testing: scanning source code for security flaws.
- Scaffold
- To set up the basic structure of a new software project.
- ScrollRestoration
- A React Router component that restores the scroll position when pages change.
- Self-distillation
- A model learning from its own explanations.
- Semantic search
- Search that matches by meaning instead of exact words.
- Shearing layers
- Frank Duffy's term for concerns that change at different time scales.
- Skill
- A saved set of instructions that an AI agent can follow.
- Slop
- Low-quality output produced carelessly.
- Slopotype
- The essay's word for a prototype made quickly by AI with little thought.
- SOC 2
- An audit standard that many companies read as requiring human code review.
- Spike
- A quick experimental attempt at a change, often thrown away.
- SRE
- Site reliability engineer, a person who keeps production systems running.
- Stochastic
- Involving chance, so the same input can give different output.
- Strangler fig migration
- A rewrite that replaces an old system piece by piece.
- Structured wiki
- A wiki whose pages hold typed, machine-readable data.
- Sub-agent
- A smaller AI agent run by a main agent to do one job.
- SvelteKit
- A web framework built on the Svelte interface library.
- T-shaped engineer
- A person with deep skill in one or two areas and broad familiarity with many.
- Tab-complete
- An editor feature that suggests the next code and accepts it with one key.
- Tech debt
- Future cost created by shortcuts in code.
- Technical debt
- The future cost of shortcuts taken in code today.
- Time-to-PR
- The time from starting a piece of work to opening a pull request for it.
- Transpiled JavaScript
- JavaScript produced automatically by converting code written in another language.
- Triager
- A cheap first-pass agent that decides whether a problem needs deeper work.
- UI prototyping
- Making rough working versions of user interface ideas.
- UMAP and t-SNE
- Two methods for plotting high-dimensional data in two dimensions.
- Vibe coding
- Building software by describing what is wanted to an AI model.
- Well-founded recursion
- Recursion where each call gets closer to an end, so it must stop.
- Working memory
- The mental workspace used to combine new and stored information.