Every Engineer Is a Tech Lead Manager

The software industry has a role called Tech Lead Manager — senior enough to run a team, still close enough to the code to build. It was always a narrow, awkward rung: half your day in the work, half of it spent multiplying other people. Claude Code made that rung the default. Hand it a whole project instead of a task and it writes a plan, splits the plan, and runs copies of itself against the parts — a lead reading the codebase while workers build in parallel. You didn't write any of it. You managed it.

The clearest version of this is Claude Agent Teams. Turn it on and your session becomes a team lead: you describe the roles you want in plain English — "spawn three teammates, one on the API, one on the schema, one playing devil's advocate" — and Claude spawns them as separate instances, each with its own context window. They coordinate through a shared task list they claim work from, and a mailbox that lets them message each other directly instead of routing everything through you. You can even require a teammate to submit a plan for approval before it's allowed to touch code. That isn't autocomplete. That's a standup.

Around that sits a kit you can teach yourself in an afternoon. A persistent brief in CLAUDE.md so every teammate knows the house rules; invocable skills and slash commands for the moves you repeat; reusable subagent definitions so "security-reviewer" means the same thing every time. Then the community layer: plugin marketplaces of pre-built specialists like wshobson's agents or VoltAgent's hundred-plus subagents, and heavier orchestrators like Ruflo, the framework that grew past sixty agents before Anthropic asked it to change its name. Codex CLI, the other tool I live in, has the same primitives. Assembling a team is now a config file, not a company.

It works because a single agent hits a wall that has nothing to do with intelligence. Context is finite. Point one model at a large codebase and it drowns — forgetting the top of the file by the bottom, blurring concerns, losing the thread. Give each teammate one narrow job in a clean context and the drowning stops. Three focused agents in three clean contexts beat one generalist holding the whole system in its head. Isolation, not horsepower, is the trick.

None of it is free. A fleet burns tokens — Anthropic's own research agents ran through roughly fifteen times what a normal chat costs — and it fails in a specific way: small mistakes compound. Write code by hand and the pain is immediate, so you fix as you go. Turn a team loose and a dozen harmless-looking errors drift in a dozen directions before you look up.

Which is the oldest lesson in management. Generating code was never the bottleneck; deciding whether it's right always was — and a team of agents makes the first part nearly free while handing the second entirely to you. So the manager's skills are the ones that pay now: a spec precise enough to survive being copied across a fleet, a team kept small enough to actually review, and the taste to know what to keep. You're not writing the code anymore; you're running the agents that write it. And every engineer is levelled up.