AI-Assisted Software Development

AI-assisted software development is a human-in-the-loop practice where LLMs and coding agents help research codebases, draft plans, write tests, implement changes, debug failures, and generate support tooling while the human supplies intent, taste, constraints, review, and final responsibility.^{source: mark-erikson-ai-thoughts-part-1-2026.md}

Mark Erikson's "My Thoughts on AI" describes a shift from fear and refusal to pragmatic use: first asking AI to explain unfamiliar architecture, then using it to write targeted tests, attempt library features, build lint rules, optimize Immer, and expand Replay/React instrumentation. The common pattern is not "turn the brain off forever," but using agents to accelerate work while repeatedly rebuilding a mental model and correcting the output.^{source: mark-erikson-ai-thoughts-part-1-2026.md}

The article's most useful technical stance is that non-deterministic AI output can be made sufficiently bounded by deterministic scaffolding: tests, typechecking, linting, CI, code formatters, static analysis, prompt/context files, explicit plans, and human review. Erikson still wants deterministic code and predictable systems; he argues the trick is to minimize what the LLM must improvise, encode repeatable knowledge into scripts and tools, and use the LLM around that automation.^{source: mark-erikson-ai-thoughts-part-1-2026.md}

This connects directly to harness-engineering: agent productivity depends on the surrounding harness, not just model capability. It also complicates personal-agents because the best experience may feel like delegated work, but it still relies on context, permissions, diagnostics, and verification loops.

Addy Osmani's cognitive-surrender framing adds a sharper failure mode: AI help becomes dangerous when generated output replaces the engineer's independent view instead of extending it. In this view, the same tool can either build skill through conceptual inquiry and review, or create comprehension debt when code ships faster than human understanding grows.^{source: addy-osmani-cognitive-surrender-2026.md}

Thariq's html-artifacts proposal adds an interface layer to this practice: if agents are producing large specs, reviews, reports, and plans, the output format should help humans understand and steer the work. HTML artifacts can render diagrams, annotated diffs, prototypes, and custom editors, making review and iteration easier than reading a long markdown file.^{source: thariq-unreasonable-effectiveness-html-2026.md}

Open questions the source raises:

how juniors build expertise when agents can produce plausible code before the human has a deep mental model;
how teams scale review, QA, tests, deployment, and planning when code generation becomes cheap and fast;
where the sustainable limit is for parallel agent sessions and AI-driven shipping speed;
whether software craft shifts from hand-writing code toward supplying context, taste, architecture, guardrails, and judgment.

The speedup-illusion paper cautions that felt ease is not the same as speed. In a preregistered study of short cognitive tasks, AI reduced subjective effort but did not reliably reduce elapsed completion time. For software work, this means productivity claims need to include prompting, reading, verification, and integration time, not just the model's answer latency.^{source: yu-cognitive-offloading-speedup-illusion-2026.md}

Maguire's ai-job-grief essay adds the identity layer: when AI-assisted development changes what counts as valuable expertise, workers may experience not only anxiety about job loss but grief over the devaluation of professional judgment. This is especially relevant to juniors and generalists whose career paths depend on accumulating exactly the judgment that agents appear to compress.^{source: jack-maguire-ai-job-grief-2026.md}

Christoph Nakazawa's modern-engineering-values essay gives a more agent-native endpoint of the same practice. He describes rarely writing code by hand, running Codex CLI across multiple projects, asking agents to gather context and propose plans before execution, forcing bug fixes to start with failing tests, and treating human attention as the scarce resource for discussion, validation, review, and judgment rather than for typing code.^{source: matt-van-horn-wtf-is-a-loop-2026.md}

Milan Jovanovic's dry-principle essay adds a non-AI software-design constraint that becomes more important when agents make refactoring cheap: similar-looking code is not necessarily duplicated knowledge, and premature extraction can create hidden coupling across module boundaries. Agentic development therefore needs semantic review of abstractions, not just tests and fast code generation.^{source: milan-jovanovic-dry-misunderstood-2026.md}

At Shopify scale, shopify-river shows AI-assisted development as a social system rather than a private IDE loop: agent-created PRs are authored by River but reviewed by humans, and the surrounding Slack conversations let employees observe how others scope requests, debug, query logs, and improve shared instructions.^{source: tobi-lutke-learning-shop-floor-river-2026.md}

Van Horn's agent-loops article describes the next abstraction step for AI-assisted development: rather than prompting a coding agent directly, the engineer writes a loop that prompts agents, inspects results, verifies progress, caps cost, and keeps going on infrastructure time. Osmani's "Loop Engineering" adds the productized building blocks now appearing in Codex and Claude Code: automations, worktrees, skills, plugins/connectors, sub-agents, and durable memory. This intensifies the page's existing open question about scaling parallel agents, because the bottleneck becomes loop management, verification, review bandwidth, and stopping conditions rather than typing code.^{source: addy-osmani-loop-engineering-2026.md}

Armin Ronacher adds a more skeptical loop-era position. He accepts that loops already work well for porting, benchmarks, security scanning, and exploratory research, but argues that long-lived code is different: present models tend to add local defenses, weak invariants, duplicate logic, and abstractions that may pass tests while degrading human comprehension. The pressure to use loops remains real, because attackers, security reporters, and competitors can loop even if a maintainer tries to opt out.^{source: armin-ronacher-the-coming-loop-2026.md}

AI-Assisted Software Development

Resources