Latest posts
-
Verifying work you did not write and cannot ask about
An agent that can run your tests can check its own work. An agent that writes the tests after the code checks its own assumptions instead, which is worthless. Here is the ordering that fixes it and what a verification gate should actually contain.
-
Giving an agent access without giving it everything
An agent with a shell has whatever your shell has. Approval prompts stop being a control the moment there are too many of them. Here is the threat model that actually applies, and the four controls that survive contact with daily use.
-
MCP explained: what it does, and what it does not
MCP is being sold as the thing that makes agents useful. It is narrower and more valuable than that: a standard way for a tool to describe itself to a model. Here is what that buys you, and the problem it very clearly does not solve.
-
Reviewing code your agent wrote is a different job
Normal review assumes an author who can explain their reasoning and who wrote at human speed. Neither holds. Here is what changes: the defects cluster differently, the volume breaks the process, and plausibility stops being evidence.
-
Writing an instruction file your agent will actually follow
AGENTS.md and CLAUDE.md are the highest-leverage files in a repository using agents, and most of them are written wrong. Long, aspirational, full of things the agent can already see. Here is what belongs in one and what a 400-line file costs you.
-
Claude Code vs Cursor: two different bets about where you work
This comparison is usually run as a feature fight and it is not one. The two tools made opposite bets about where a developer sits and how much they watch. Here is what each bet buys, what it costs, and which one describes you.
-
Why your coding agent forgets, and what to actually do about it
An agent has no memory between sessions and a finite window inside one. Most complaints about agent quality are really complaints about context management. Here is what fills the window, what falls out of it, and the three fixes that survive a session ending.
-
The best AI coding agents, sorted by what they are actually for
Every ranking of coding agents is out of date within a quarter and most of them measure the wrong thing. Here is the durable version: four shapes of tool, what each is genuinely built for, and the question that tells you which one you are.
-
What an AI coding agent actually is (and how it differs from autocomplete)
Autocomplete predicts your next line. A chat assistant answers your question. An agent decides what to do, does it, looks at the result and tries again. That third thing is a different category, and the difference shows up in your repository.