#agentic-coding
Wiki 40
- 1Password — What We Learned Using AI Agents to Refactor a Monolith 1Password decomposes a multi-million-line Go monolith with AI agents; the win is agent-built deterministic tools, not agent-written code
- A Voice From Nowhere zanlib on Leo XIV's encyclical: generated work copies the form of thought, not the path under it
- Agent-built deterministic tools Use the LLM to build deterministic analyzers and manifests once, then constrain all subsequent agent work to those stable outputs
- Agents as first-class identities Giving agents their own key pairs and signed event trails instead of running them through a shared bot account — attribution survives, but effort still doesn't
- AI Didn't Make Programming Easier. It Just Made It Differently Difficult Osborn's CACM piece grounding the AI-coding shift in memory research — the burden moves from recall to judgment, and judgment is harder to build than recall was
- Benchmarking Opus 5 on SlopCodeBench Dex Horthy runs three Claude models through SlopCodeBench; the best, Opus 5, clears 24% strict pass
- Block, Inc. Jack Dorsey's payments company (Square, Cash App), also the publisher of Goose and Buzz — open-source agent tooling built on open protocols rather than vendor platforms
- Buzz: Block's Nostr-Backed Workspace for Humans and Agents Dorsey's open-source Slack/GitHub replacement gives agents the same signed identity as employees — and the runtimewire writeup checks the decentralization claim against the architecture doc
- Claude Is Not a Compiler Bleecher Snyder retires the compiler analogy — an agent's value is working vertically across layers, demonstrated by building exe.dev's distributed DNS server in a week
- Code Mode Token Savings One script against 26 tool calls on agent-swarm's own production data — a self-measured 99.2% cut
- Code Review as a Principal-Agent Problem Why review-then-commit relied on cheaply-inferable effort signals from contributors; how agents collapse that signal and what it means for review processes
- Code review throughput limits The empirical caps on effective code review (~400 LOC/h, ~1h sessions, no parallelism) and why they put a hard ceiling on "just review the AI's output"
- Credibility as the slop test Whether AI-assisted text is slop can't be judged from its surface — the only working test is whether the author would read it aloud, unashamed, staking their credibility
- Cross-model code review Who should check an LLM's code, and why the model that wrote it reviews poorly, why the reviewer must be able to solve the task, and when a weaker writer pays
- Dan Luu (danluu.com) Dan Luu's blog — measurement-driven engineering writing from a CPU-verification background, covering testing, reliability, org dynamics, and now AI coding
- Dan Luu on AI Coding, Testing, and Variance Dan Luu applies a CPU-verification background to agentic coding — testing-heavy no-review workflows, the meaninglessness of single-number model benchmarks, and working around agent failure modes
- David Crawshaw (crawshaw.io) David Crawshaw's blog at crawshaw.io — co-founder of exe.dev, ex-Tailscale, ex-Google; writes on agent-driven engineering, small teams, and infrastructure
- Differential spec analysis Build the same system with several independent agent loops, diff the finished implementations, and treat every divergence as a decision nobody specified
- exe.dev blog (blog.exe.dev) Company blog of exe.dev, the nine-person fast-booting-VM startup; agent-driven engineering practice written up from inside a team that ships on it
- How Do We Stop Vibe Coding? Alex Klos surveys spec tooling and argues vibe coding needs trust machinery rather than discipline
- Leverage Code Review for Sustainable AI Coding Development CACM news piece where practitioners say review of AI code shifts from end gate to continuous risk check aimed at assumptions, trust boundaries, evidence
- LLM output variance Run-to-run and task-to-task variance in LLM/agent output is high enough that small-sample comparisons and single-number benchmarks support almost any conclusion
- Maybe We Shouldn't Be Reviewing All This Code Rachel Laycock on moving review's jobs earlier into pairing, design sessions and automation, keeping human code review for the exceptions
- Memorizing Session Transcripts Isn't Useful theahura found zero SWE gain from agents searching their own past transcripts, and explains why
- Portal by Spotify Cut Claude Code Token Usage by 90% Spotify PM routes Claude Code's bulk file reads and boilerplate to a cheaper model via hooks; the 90% covers bulk reads, not total usage
- READMENOT — a marker for code not meant for humans William Woodruff's proposal for a READMENOT file that honestly signals a project is slop — not meant for unwitting human reading
- Recall-to-judgment shift AI moves programming's cognitive burden from retrieving syntax and patterns to evaluating whether generated output is correct and appropriate — relocation, not reduction
- Reviewing AI Code Thomas Depierre's empirical case that "just review the AI's code like an intern's" collapses against what research says about code-review throughput and reviewer overconfidence
- Slop-marker convention The idea of an explicit, machine-and-human-readable marker declaring code (or text) as slop — defined by its production process, not its surface
- Starling Desktop A complete Linux desktop claimed to be AI-written in six months, weighed as evidence
- Testing-heavy, no-review workflow Ship large volumes of unreviewed (including agent-generated) code by trusting randomized testing over human code review — the CPU-verification model applied to software
- The Agent Principal-Agent Problem Crawshaw on how agents broke code review by collapsing the effort signal between contributor and reviewer; the small-team escape and the big-co dead end
- The LLM Critics Are Right. I Use LLMs Anyway Jeremy Theocharis on the dissonance of agreeing with nearly every anti-LLM argument while spending $10k/month on tokens — and the workflow patterns that make it defensible
- The Short Leash AI Coding Method Read every diff in the permission prompt and deny freely — Slepak's method for critical software
- theocharis.dev Jeremy Theocharis's blog on LLM-assisted work, local-first software, and engineering practice; writes with LLMs about using LLMs
- Thoughts on Coding Agents (rakyll) Agents amplify direction rather than supply it, which removes the permission tax large orgs charge
- Vibe-engineering Using an agent to accelerate your own decision-making across every layer of a system, rather than delegating the decisions — the line that separates it from vibe-coding
- Why Software Factories Fail Dex Horthy on why lights-off factories fail — RL rewards passing tests, never good design
- Working around agent failure modes is the skill Most of the skill in using coding agents is knowing their specific failure modes and building around them — a skill that decays as labs fix those failures
- Your AI Agent Doesn't Understand Code, It Guesses Confidently Vendor post on CodeSlicer, an impact graph that separates proved call edges from plausible guesses
Toolbox 6
- agent-shell Native Emacs mode for ACP-speaking coding agents, with one buffer interface across every agent
- Boffin Routes per-edit architectural constraints into a coding agent and requires a proportional check
- mindwalk Replays a coding-agent session as light moving over a 3D map of the repository it touched
- OneCLI Credential gateway that swaps an agent's placeholder key for the real secret on the way out
- re_gent Git-shaped version control for AI agent activity, with log, blame and show per tool call
- Toolcraft Open-source React starter that hands an AI agent a design-app architecture, not just components
talks 1
- Software Fundamentals Matter More Than Ever Matt Pocock argues AI coding rewards old fundamentals — shared design concept, ubiquitous language, TDD, deep modules — and bad code costs more than ever