The Year of the Agent Swarm
Steve Yegge published a chart of agentic development stages. It starts at "zero AI" and climbs through IDE copilots, YOLO mode, CLI agents, and eventually to building your own orchestrator. Eight stages. I put myself barely in stage seven — multiple parallel instances, concurrent worktrees, and I've already got Claude running in a custom orchestrator mode by default. It delegates implementation to subagents, preserves context aggressively, enforces file ownership boundaries. Homebrew orchestration.
But it's not Gas Town scale. Not yet.
Gas Town is Steve's answer to stage eight: an orchestration layer where a "Mayor" agent coordinates worker "Polecats" across isolated git worktrees, tracking work through Beads and handling the chaos of twenty or thirty agents at once. His warning is explicit: if you're not at stage seven, you're not ready. The chimps will wreck your shit.
I believe him. I'm close, but I'm not there yet.
But I can see it coming. 2026 is shaping up to be the year agent orchestration goes mainstream. The primitives are crystallizing: durable task tracking that persists beyond a single session, and reliable messaging between agents so they can coordinate, review each other's work, hand off context.
Beads takes the git-native approach — tasks stored in .beads/ inside the repo, using git itself for coordination across branches and machines. Claude Code's new Tasks system (which Anthropic explicitly says was inspired by Beads) goes filesystem-native instead — persisted to ~/.claude/tasks, broadcast across sessions sharing a Task List. Different trade-offs, same insight: agents need durable state that survives context boundaries.
Other approaches are converging on similar insights. The Ralph Loop takes multiple fresh passes over a task with clean context each time — different mechanism, same underlying bet that iteration and verification beat single-shot heroics.
Meanwhile, swarm mode is emerging in Claude Code's feature flags — native multi-agent coordination with teammate messaging and task delegation. That's the one I'm watching. If Anthropic ships it properly, it might be the on-ramp I need to full-stage seven without hand-rolling a true orchestrator.
Here's my honest position: I'm stretched thin. Five Claude projects at a time, two of them with concurrent worktrees. For games I still spend too much time as the human in the loop — the creative direction can't be fully delegated yet. For work, the discovery phase still needs me present. I could probably let agents cross-talk more, handle more of the boilerplate iteration, move faster at higher cost. But I haven't taken that plunge.
What I have done: automated the pieces I understand. A five-stage dialectical refinement process for specs. Periodic multi-agent code reviews at major milestones. Context-preserving implementation loops with feature reviews. Quality gates. Test coverage. But it's still too manual. I kick off each step myself, every session.
That's where I need to level up in 2026: automating the full SDLC while keeping the human in the loop alive. Not removing myself from the process — just moving to a higher altitude.
I suspect if I took the Gas Town or swarm plunge, I'd be pleasantly surprised at how much I could still steer while gaining throughput. But for now I'm waiting for swarm mode to exit feature flags. Hopefully sooner than later.
The chimps are learning to coordinate. I should probably keep up.