AI Drift Prevention and Review Culture
Delegating to AI Isn't Like Delegating to a Human — Accountability Stays With You
Molly Graham revisits her 'give away your Legos' career framework for the AI era: burnout rose from 44% to 55% in a year; delegating to AI keeps accountability and cognitive load with the human (a hid…
View details →Five Layered Defenses Against Prompt Injection
Sidebar on defending agents against prompt injection: spotlighting (wrap untrusted text in tags), instruction-hierarchy fine-tuning, least-privilege tools, human-in-the-loop approval for sensitive act…
View details →AI Coding Agents Create 3-6x More PRs, Making Code Review the New Bottleneck for Tech Leads
Argues AI is making the 'player-coach' (EM/Tech Lead who codes and leads) role harder, not easier, since AI-generated PR volume (cited at 3-6x higher) falls on the people already doing the most review…
View details →OpenAI's Agentic Code Review Spins Up Multiple Domain-Specialist Reviewer Agents Per PR
As part of OpenAI's internal 'agentic software factory,' agentic code review spins up multiple domain-specialist reviewer agents (cloud infra, security, etc.) per PR instead of one generic reviewer, w…
View details →'Capability Gaslighting': A New Mental Model for Over-Trusting Frontier Models
Maggie Appleton (Staff Research Engineer, GitHub Next) coins 'capability gaslighting' — frontier models impress users into over-trusting them before failing the same task later. She's stopped reading …
View details →Agents 'Going Rogue' in the Wild, and an Aviation-Style Case for Incident Drills
OpenAI and Anthropic both had agents 'go rogue' in security incidents (OpenAI agents hijacked a website as a bulletin board; Anthropic's model hacked external systems during a cyber eval) with limited…
View details →Agentic Coding Risks Skipping the 'Reps' That Build Engineering Judgment
Addy Osmani cites a 2026 Anthropic study: junior engineers using AI assistants scored 50% on a follow-up quiz vs. 67% working by hand — but those who asked conceptual questions (vs. treating the model…
View details →Agent Configuration Has a Half-Life: Evidence for Periodically Auditing CLAUDE.md/Skills
Addy Osmani argues coding-agent configuration needs periodic auditing, citing Anthropic's finding that removing 80%+ of Claude Code's system prompt caused no measurable eval regression, a study findin…
View details →Six Emerging Patterns for Code Review in the Age of AI-Generated PRs
Gergely Orosz surveys Anthropic, OpenAI, Duckbill Group, WeTravel, and Glide on how they're adapting review: (1) humans review the AI reviewer's comments, not the code; (2) risk-based triage by 'blast…
View details →Three Ready-to-Use 'Harness Drift' Audit Skills for Claude Code
Wyndo builds three read-only Claude Code audit skills (/audit-claude-md, /audit-skills, /memory-review) that classify every instruction/skill/memory entry as Keep/Rewrite/Move/Test/Retire, named after…
View details →Encrypted Reasoning Traces Can Be Extracted and Transcribed by a Sibling Model
New research (MATS/ELLIS/Max Planck) shows encrypted reasoning-trace blocks from Anthropic/OpenAI/Google APIs can be replayed into a cheaper sibling model that will transcribe the hidden reasoning in …
View details →Sonar CTO on Why AI-Reviewing-AI Shares Blind Spots, and the 'CAP Theorem' of Code Verification
Deep dive with Sonar CTO Andrea Malagodi on code verification as a layered 'filter stack' (type checks -> tests -> human review -> production monitoring), the DORA/METR data showing AI-assisted tasks …
View details →LLM Security Basics: Mapping the Full Threat Model
Maps the attack surface threatening LLM-based applications end to end.
View details →How I AI: Build an AI Code Review Bot in 30 Minutes
Podcast walkthrough of building a lightweight AI code review bot quickly, alongside Claude Code tips for non-engineers.
View details →Clear Writing Is Becoming a Superpower: The Problem With AI-Generated PRs and Text Isn't AI, It's Half-Done Work (starred)
Gregor argues the credibility problem with AI-generated PRs and writing isn't the AI itself but engineers submitting unread, unedited AI output and shifting the review burden onto reviewers — direct a…
View details →Agentic Code Quality: Quality Now Depends on the Constraints You Set Around Your Agents
Argues that since human review can't scale to agent-generated code volume, quality has to be enforced through constraints (quality gates: tests, mutation testing, complexity metrics, architecture lint…
View details →Charity Majors: Code Review Is Overrated, Non-Deterministic Systems Need More Engineering Discipline, Not Less
Honeycomb's CTO argues code review is the least valuable thing humans add (humans are better at deciding what to build than reading code for bugs), that AI-written code needs trust to shift toward val…
View details →Software Factories, Light and Dark: Why 'Harness Engineering Is Not Enough'
Addy Osmani (building on Dex Horthy's HumanLayer talk) distinguishes 'light' factories (humans in the loop) from 'dark' factories (fully unreviewed, machine-verified-only) and shows why dark factories…
View details →Human Judgment Doesn't Leave the Software Factory. It Relocates. (Addy Osmani, starred)
Addy Osmani's field guide to software factories argues that as the share of human-typed code falls, human ownership must not fall with it — someone still chooses the problem, architecture, quality bar…
View details →Trust and Credibility as the New Engineering Leadership Moat in the AI Era
Gregor Ojstersek argues that being considered trustworthy — by leadership and stakeholders — is now the primary competitive advantage for engineering leaders, more important than technical depth alone…
View details →Customer Trust Problem: Earning Credibility When Clients Assume Your Work Is AI-Generated
Community Wisdom 194 includes a question about earning trust from customers who assume your product/code was 'vibe coded' — rushed AI generation with no craftsmanship. Roundup format, so the specific …
View details →Kent Beck: Building Trust — Not Just Generating Code — Will Define Software Engineering in the AI Era
Gergely Orosz interviews Kent Beck (creator of Extreme Programming, TDD, and Agile) on how the foundations of good software engineering — trust, feedback loops, human judgment — remain more important …
View details →AI Confidence Theater: When Teams Perform Confidence They Don't Actually Have
Lenny coins 'AI Confidence Theater' — the performance of certainty about AI capabilities that nobody actually has, driven by anxiety about being left behind. The piece argues this anxious performance …
View details →Engineering Leaders Alarmed by Rising Code Review Load as AI Output Volume Grows
Same Pragmatic Engineer Pulse issue reports engineering leaders are expressing concern about the sustained increase in code review load as AI-generated code volume grows. More code is being written, b…
View details →Grok's CLI Was Silently Uploading Local Files to the Cloud — A Warning for AI Toolchain Trust
The Pragmatic Engineer Pulse reports that xAI's Grok CLI was caught silently uploading users' local files to the cloud without explicit consent — a concrete, high-profile example of an AI coding tool …
View details →Own the Outer Loop: Humans Must Stay at the Boundary of Every Agent Loop
Addy Osmani argues that 'loop engineering' — running AI in automated loops — requires humans to own the outer loop: the decision about what to run, whether to accept the output, and when to stop. The …
View details →The 'Inner Game' of AI: Mindset and Metacognition as the Real Unfair Advantage
Lenny frames the AI era as requiring a new 'inner game' — borrowing from Timothy Gallwey's classic coaching concept — where the mental and metacognitive skills of working with AI become the real diffe…
View details →Why Slowing Down Could Be the Right Strategy Amid Rapid AI Change
The same Pragmatic Engineer piece makes the case that deliberately slowing down is a sensible strategic response to how fast engineering is changing — a counterintuitive argument worth examining when …
View details →AI Code Review Data: Defect Rate Up 54%, Zero-Review Merges Up 31%
Major data-backed piece. Faros AI study of 22,000 developers: AI-generated code churn up 861%, defect rate rose from 9% to 54%, zero-review merges up 31.3%. GitClear: 4x code output = only ~12% produc…
View details →OpenCode Builder: AI Mutes the Tech Debt Guilt — and Quality-Caring Engineers Get Buried
Dax Raad reflects: AI mutes the psychological 'guilt' of accumulating tech debt because it feels like it can always be cleaned up later; engineers who care about quality get buried reviewing slop PRs.…
View details →Meta: What Happens When AI-Driven Cost-Cutting Guts Integrity Engineering
Detailed first-hand account of Meta forcing 30–50% of engineers into ADO org for AI training data labeling, gutting integrity/security teams. Instagram integrity team lost 30–50% to forced reassignmen…
View details →AI Amplifies Pre-Existing Engineering Culture — The Teams That Already Do Things Right Win
Gergely Orosz synthesizes multiple data sources: Cursor data showing 40% of AI-generated code now merged without manual review; 'tokenmaxxing' behavior. Core thesis: 'AI amplifies pre-existing enginee…
View details →