This post describes the change in how developers work with AI as a move from “completion” to “delegation”: instead of suggesting the next lines in an editor, a coding agent now takes an instruction, makes a plan, reads the relevant files, edits several of them, runs the tests and fixes its own failures. It frames the resulting practice as Agentic Coding and argues that the human role is shifting from writing code to setting direction and reviewing what the agent produces.
The bulk of the post is three principles for delegating implementation work, followed by two areas the author treats as becoming more important rather than less once an agent writes most of the code: reviewing its diffs, and limiting what it is permitted to do.
Key Points
- The post proposes viewing an agent as “model plus harness”: the same model can be much more or less useful depending on the surrounding machinery for file operations, test execution and permission management (see Agent Harness).
- Its survey of tools as of 2026 — Claude Code, OpenAI Codex, Cursor, GitHub Copilot and Cline — concludes that each has trade-offs and that the practical answer is to choose by workflow.
- First principle: split work into pieces a human can review. Delegating large tasks feels faster, the post argues, but a diff too large to review ends up as rework.
- Second principle: give the agent a means of verification before it starts, such as the test command to pass, so that it can judge success itself and iterate; the author states that “build, let the agent test itself, then review” runs far faster than “build, have a human try it, then ask for fixes”.
- Third principle: provide context through files in the repository. The post describes AGENTS.md as a tool-independent convention read by several agents and CLAUDE.md as the file Claude Code reads, and recommends recording project layout, conventions, commands and prohibitions there; it notes that layering these files from home directory to project root to subdirectory is common practice.
- Because agents generate code quickly and in volume, the post treats reading every diff as the core of quality: rejecting unintended or excessive changes, asking why an implementation was chosen rather than accepting it because it runs, and keeping tasks small on the assumption that review will become the bottleneck (compare Review Bottleneck).
- On security it recommends least privilege with permission required for destructive commands, keeping API keys and passwords out of prompts and repositories, and instructing the agent not to follow instructions found in external documents or web pages it reads, as a precaution against Prompt injection.
Context
The post is a general, introductory overview written from the author’s own practice rather than a report of measured results; its recommendations are presented as an individual practitioner’s summary of the state of things in 2026, and its tool characterizations are one-line impressions.