Every job posting I see right now lists the same five tools. Claude Code. Cursor. MCP. Some flavor of agent framework. As if hiring for typists in 1985 by listing brands of typewriter.
The variance in shipped output isn't coming from tool choice. It's coming from people who can sit in front of a codebase nobody has touched in eighteen months and, in an afternoon, tell you which three files are load-bearing, which abstraction is the team's pet, and which migration is silently keeping someone up at night.
What codebase fluency actually is
It's pattern recognition built over years of being wrong. You learn that a folder called `utils/` with 47 files is a confession, not a structure. You read a commit history and notice the three weeks where commits stopped. That's where the rewrite was attempted and abandoned.
- Reading a `package.json` and inferring the team's last six months of priorities.
- Spotting which tests are theater and which are actually running in CI.
- Recognizing when a class is a noun pretending to be a verb.
- Knowing which kind of `// TODO` is a real promise and which is a tombstone.
Why this matters more, not less, with agents
An agent can produce a thousand lines of plausible code in two minutes. The cost of plausible-but-wrong has dropped to near zero on the producer side and stayed exactly the same on the consumer side. Someone still has to read it, route it, and own it when it ships.
Agents amplify clarity, and they amplify confusion at the same rate. The differential is the engineer.
If you're hiring in 2026 and your interview loop doesn't include 'sit with this real codebase for ninety minutes and tell us three things you'd refuse to ship', you are hiring for the wrong shape of person.