defract › blog

writing

field notes on running parallel Claude Code agents, the engineering lifecycle, and owning a product end to end.

2026-07-31 6 min read

the acceleration whiplash: what AI actually did to engineering teams

A 2026 Faros study measured AI's real impact on engineering orgs: shipping up ~66%, but incidents per PR up ~243% and review time up ~441%. The speed is real — so is the damage. The data, and what it means.

2026-07-30 6 min read

reviewing AI code one task at a time isn't enough

task-level review is mandatory. without a codebase-level strategy across tasks, you can ship a hundred correct changes and still end up with a broken system.

2026-07-30 6 min

the last mile: why AI coding agents stop at PR open

PR open is not PR merged. The gap between them — CI failures, rebase conflicts, duplicate PRs, stuck queues — is where agent-assisted development loses its gains. Here is how to close it.

2026-07-28 7 min read

When does your AI coding agent's context run out?

AI coding agents don't announce when their context is full — quality quietly degrades. Here's how to read the signals, when to compact, and what it means when you're running agents in parallel.

2026-07-23 5 min read

AI code review that fixes instead of flagging

Most code review findings are small and obvious. Here is when an AI reviewer should apply the fix inline — and when it must loop back to implementation.

2026-07-22 6 min read

the year of harnesses

The frontier models converged, so the differentiation moved up a layer — to the harness: how work is decomposed, where a run pauses for a human, and what gets checked before code ships. On agent harnesses and harness engineering.

2026-07-22 8 min read

which agentic coding patterns actually scale

Most agentic coding patterns work on the first feature. Here is what changes as the codebase grows, the task complexity rises, or you try to share the loop with anyone else — and which patterns hold up.

2026-07-21 7 min read

getting from Step 1 to Step 2 of AI coding adoption

Most engineers using Claude Code stop at Step 1: one agent, watching every line. Here is the recipe for Step 2 — parallel agents on worktrees, self-verification loops, automated review, and the trust you build to stop reading everything.

2026-07-16 5 min read

how to make AI agents follow your CLAUDE.md project conventions

Every parallel Claude Code agent starts fresh. Here is how to get your project conventions, architecture decisions, and prohibited patterns into every agent prompt without restating them per task.

2026-07-15 7 min read

what to do with everything your AI coding agent flags

Your coding agent surfaces a stream of suggestions — tech debt, edge cases, follow-up tasks. Here are the five patterns developers actually use to handle them without building a graveyard.

2026-07-14 7 min

when AI coding agents scope before they're ready

AI coding agents default to confident scoping even when they're missing key context. Here is what premature commitment looks like and three mechanics that change the dynamic.

2026-07-09 8 min read

configuring Claude Code for production: hooks and permissions

Most teams run Claude Code with either constant permission prompts or --dangerously-skip-permissions. There is a middle path: a settings.json with typed hooks, allow lists, and deny lists. Here is what ours looks like.

2026-07-08 7 min read

Claude Code worktrees: parallel agents without the conflicts

Running multiple Claude Code agents at once is fine until two of them edit the same file. Git worktrees are the fix — here is the practical workflow, and where the overhead still lands on you.

2026-07-07 6 min read

when your AI coding agent proposes the next task

Most AI coding agents only execute what you assign. Real development generates adjacent work mid-task. Here is how agents can surface that work for your review without going off-script.

2026-07-02 6 min

why AI coding agents need hard stage boundaries

Two failure modes break multi-agent Claude Code workflows: stages that advance automatically, and agents that do out-of-stage work. Here is how both happen and why you need enforcement at two separate levels.

2026-07-01 8 min read

claude code skills frameworks: Superpowers vs GSD vs gstack

What Superpowers, GSD and gstack each constrain, why experienced users stack them rather than pick one, and when you'd reach for an integrated app instead.

2026-06-30 7 min read

from prompt to production: a structured pipeline for Claude Code

Running Claude Code on real features needs more than a prompt. A structured workflow — scope, design, architecture, implementation, review, release — is what keeps production output coherent.

2026-06-24 9 min read

best claude code orchestrators in 2026: an honest roundup

an honest, vendor-neutral roundup of Claude Code orchestrators in 2026 — parallel runners, spec-first tools, and the native floor — with where each one genuinely wins.

2026-06-24 7 min read

local-first ai coding: why your transcripts should stay on your machine

what local-first means for AI coding tools, the data trail an agent session leaves, who you actually pay for inference, and a vendor-neutral checklist to verify the claim.

2026-06-18 6 min

spec-driven development, explained

spec-driven development means writing the spec - requirements, design, tasks - and correcting it before agents write code. what it is, where it helps, and its honest limits.

2026-06-18 7 min read

why 20 Claude Code instances break down (and what to do)

scaling Claude Code from 5 to 20 parallel agents crosses a threshold where qualitatively different failure modes appear — rate limits, supervision collapse, context cascade failures. what breaks, and what to do instead.

2026-06-17 6 min read

compound engineering with Claude Code: the loop, and the step everyone skips

Compound engineering — each task makes the next one easier — is the real shift in AI coding. Here's the loop, why the 'compound' step is the one everyone skips, and how to make it structural.

2026-06-17 7 min read

keeping context and decisions consistent across parallel AI agents

run several Claude Code agents at once and they drift — each has its own context window, none remembers what the others decided. what actually needs to be shared, and how teams keep parallel agents coherent.

2026-06-17 6 min read

AI didn't kill software craft. it moved it.

the worry that AI killed craftsmanship points at real grief and the wrong conclusion. craft didn't die — it moved upstream to taste and selection, and downstream to review.

2026-06-17 6 min read

you can build anything now. that's the new bottleneck.

when implementation gets cheap, the constraint moves to the front: deciding what's worth building. the scarce skill is product and business judgment, not typing.

2026-06-17 6 min read

why AI coding agents agree with everything (and how to make them push back)

LLM agents are trained to be agreeable, so they rubber-stamp your design and approve their own code. why that happens, and the tactics that get real pushback.

2026-06-17 8 min read

how AI agents onboard to a legacy codebase you've never touched

drop agents into an unfamiliar codebase and they can fix real bugs fast — if you run the right workflow. map, conventions, scoped tasks, worktrees, validation, and where it fails.

2026-06-17 6 min read

giving AI agents roles: PM, architect, reviewer, QA

one generalist agent collapses scoping, building, and reviewing into a context that reviews its own work. splitting into roles — with hand-offs and an independent reviewer — does better.

2026-06-16 8 min read

the cognitive load of running parallel Claude Code agents

running multiple Claude Code agents in parallel is technically possible today. this is what the cognitive overhead actually looks like — and why throughput alone doesn't solve it.