If you've been living under a rock, the new shape of AI coding is not one agent doing everything — it is one orchestrator dispatching many agents in parallel. Sequential agents are the COBOL of 2026. Parallel-with-isolation is the actual unlock, and tonight I shipped fourteen feature branches in four hours using exactly that pattern.
Spoiler: the recipe is three ingredients — a master plan doc, git worktrees, and caveman-ultra mode. Behold.
The Setup
One markdown file at docs/plans/MASTER_PLAN.md is the source of truth. It has a numbered task table and a one-line agent prompt template. Each agent reads the file, claims an unclaimed task by section ID, and ships on a feature branch named after that ID. No coordinator code, no queue server, no Redis — git is the queue.
The Money Pattern
The orchestrator dispatches with the Agent tool and the isolation:worktree flag. Every worker gets its own git worktree at .claude/worktrees/agent-HASH, which means five agents can run git checkout at the same time on the same repo without trampling each other's index. Branch collisions die.
Caveman-ultra is the budget multiplier. The mode cuts agent token output by roughly 75 percent — fragment sentences, dropped articles, abbreviated nouns, arrows for causality. Five caveman-ultra agents cost about the same as 1.5 normal agents over the same wall-clock hour. Same hourly budget, three times the throughput.
The Catch
Plot twist: caveman-ultra agents still write normal code, normal commits, and normal PR descriptions. The compression applies to chat, plans, and the under-200-word return summary — not to the artefacts. If you let an agent compress code comments or commit messages you will regret it within a week.
Also, the master plan doc needs atomic claim semantics. The first agent that flips Status to IN_PROGRESS commits that row immediately on its branch. Other agents pull, see it taken, skip. Without that handshake two agents will race the same task.
The Verdict
Parallel orchestration with worktree isolation and caveman-ultra is the actual frontier of AI coding workflows in 2026. The internet collectively lost its mind over autonomous agents, but the boring engineering answer is one orchestrator, five workers, one markdown file, three git worktrees. Tonight it shipped fourteen branches. Do not @ me about LangGraph.