X Discussion: Claude Code's Addy Osmani Says "Your Coding Agents Made You a Manager" and Maps Team Management Habits to Agent Work

Background and the pain point

Now that coding agents can run for hours, the bottleneck for developers has shifted from writing code to briefing tasks, checking results and deciding what ships. On November 8 at 07:32 UTC, Addy Osmani of Anthropic's Claude Code team published a long X article titled "Your coding agents made you a manager". His view is that engineers who spend all day with an agent are already managing, and the people who get the most from agents are not always the best prompt writers but the ones who learned to manage. At capture it had about 964 likes, 1,528 bookmarks and about 53,700 impressions; more bookmarks than likes suggests many readers saved it as a checklist. It is an employee's personal article, not an Anthropic announcement.

How it works

The article maps good manager habits to Claude Code features. To give the agent a way to check its work, you can ask it to run the tests until they pass, use /goal so a separate model checks the goal after every turn, add a Stop hook that blocks the turn from ending until the typecheck passes, or use /code-review to send the diff to a subagent that never saw the conversation. Rules come in two layers: prompts, CLAUDE.md and skills are requests the agent may weigh, while hooks, deny rules and managed settings are requirements that are always enforced. For decomposition, /batch turns a large mechanical change into 5 to 30 tasks run in separate worktrees, each with its own PR, and /rewind makes rolling back code and conversation cheap.

Benchmarks and numbers

This is a methods article with no new benchmark, and no official comparison data has been published. The only figure is the author's own account that he usually runs 4 to 5 background agents plus a few local sessions; he calls the rising attention cost past the first few agents the "orchestration tax". He suggests measuring PR cycle time, rework and reverts rather than the share of code written by AI. The claim circulating in quote posts that Claude writes about 90% of Claude Code is not in the article and is not used here.

What developers can do

Follow the checklist at the end: give every task a check it can run and ask for proof; tell the agent only what it cannot see in the repo; agree a plan in plan mode before coding; log anything you correct twice in CLAUDE.md and turn repeated workflows into skills; after two failed corrections, restart and rephrase the prompt; and review every diff as if you will be paged for it. Confirm exact commands and settings in the official Claude Code best practices docs.