When LLMs erase your expertise, token costs in agent loops, and datasette-agent-edit

Good morning. It's Sunday, June 7, and we're covering a raw first-person account of AI hollowing out a decade of engineering expertise, new research on where agent token budgets actually disappear, and Simon Willison shipping a reusable text-editor foundation for Datasette Agent.

A quieter weekend day, but the career piece hit 1,134 points on HN - the loudest conversation in the community all week.

THE BIG ONE

LLMs eroded every pillar of a 10-year engineering career

A backend engineer with a decade in finance and payment systems published a detailed, unusually honest breakdown of watching AI dismantle his competitive edge in three distinct phases. First: domain knowledge. PCI compliance, double-entry ledgers, idempotency patterns - years of accumulated specialization he thought was unreplicable. Modern models prompted with the right context matched it. Second: debugging. By mid-2025 Claude Code was one-shotting 60% of production bugs given a stack trace and Sentry MCP. By the time Claude 4.6, 4.7, and GPT 5.5 arrived with DataDog MCP, the number hit 90% - including race conditions across distributed systems that would have taken two days of human investigation. Third: architecture and code quality. Still human territory, except the industry is deliberately lowering the bar, shipping "C-grade" codebases designed for LLMs to read rather than engineers.

The post landed 1,134 points and 500+ comments on HN. What made it resonate was the structure: not a generic "AI is taking jobs" piece but a specific, phase-by-phase account of which skills eroded and when. The author does not claim to be unemployable - but argues the premium on accumulated expertise has collapsed, and that "someone has to steer the robot" is cold comfort when every senior engineer is now equally interchangeable at that task. If you work on developer tooling or are thinking about where "taste" and architecture skills fit in an agentic world, our post on skills as the new agent operating system covers the flip side of this argument.

RESEARCH

Code review eats 59% of tokens in agentic engineering pipelines

A new arXiv paper titled "Tokenomics: Quantifying Where Tokens Are Used in Agentic Software Engineering" puts hard numbers on a question every team running coding agents is quietly asking: where does all the cost go? Researchers at Concordia analyzed 30 software development tasks run through the ChatDev multi-agent framework using a GPT-5 reasoning model, mapping execution traces to standard SDLC phases - Design, Coding, Code Completion, Code Review, Testing, and Documentation.

The headline finding: the iterative Code Review stage consumed an average of 59.4% of all tokens, far outpacing initial code generation. Input tokens were the dominant type at 53.9% of total consumption, suggesting that context re-feeding across agent turns - not generation - is the primary cost driver. The practical implication is direct: if you are optimizing agentic pipelines for cost, the leverage point is not faster generation but reducing how much context gets re-sent on each review iteration. The paper proposes a methodology for benchmarking token distribution that teams can apply to their own pipelines. This connects directly to what we covered on engineering token budgets in agent harnesses.

WHAT ELSE IS HAPPENING

• datasette-agent-edit 0.1a0: Simon Willison shipped a base plugin implementing Claude's view / str_replace / insert text-editor tool pattern as a reusable foundation for Datasette Agent plugins - so every plugin that needs agentic text editing does not have to re-implement it from scratch.
• Show HN: Lathe (399 pts): A tool that uses LLMs to help you actually learn a new domain rather than skip past it - generates targeted explanations, questions your assumptions, and tracks gaps. Explicitly designed as "LLM to teach, not to do."
• How is Linear so fast? (491 pts): A technical breakdown of Linear's client-side architecture - local-first state, sync engine design, and where they trade latency. Useful reading for anyone building responsive collaborative tools.
• Anthropic, please ship Claude Desktop for Linux (534 pts): A GitHub issue thread that turned into a community demand aggregator - 500+ upvotes and counting, with Linux developers documenting workarounds and asking for parity with Mac/Windows.
• IOCCC 2025 Winners (427 pts): The 29th International Obfuscated C Code Contest results are out - perennial weekend reading for anyone who wants to remember how strange C can get when creativity is the only constraint.

FROM THE SITE

This week we published three pieces worth pairing with today's stories: GitHub Trending is full of Goose looks at Block's open-source agent dominating the trending charts, engineering token budgets in agent harnesses digs into the practical side of the Tokenomics findings above, and LLM router comparison 2026 benchmarks the current routing layer options for teams sending traffic across multiple models.

---

Every link above goes to a primary source. This brief is part of the Daily Brief archive.