superteam

🦸 Superteam

You have AI. Why are you still glued to the screen?

one /superteam command spawns a real engineering team that brainstorms the spec with you, locks executable acceptance gates, and grinds through your enterprise stack (Flyte, HDFS, k8s, internal mirrors, your MCP servers) — increment by increment, overnight, until every gate passes.
Not another agent in a loop
No fake "looks good"
No re-teaching your company every Monday

Features · Quick Start · How It Works · Run on a Remote VM · Related Projects · Global Wiki

Language / 语言: English · 中文

Blog: English · 中文

Features

Five things a single agent in a loop cannot do.

🤝 Aligned-Before-Code PM Brainstorm

"Keep asking until your goal and the spec are byte-identical."

A real PM-style intake grounded in your codebase — not a vibes-based "got it, building now." The PM surveys what you already have, asks targeted classifying questions about scope, edge cases, and integration points, and refuses to move on while ambiguity remains. No more "the LLM made an assumption I'd never make" at hour 14 of a run.

🔒 Unbribable Acceptance Gates

"You approve scripts, not adjectives. After that, 'done' is binary — and no agent can negotiate it."

Acceptance criteria compile to executable shell scripts — pytest, ssh edge-01 'hdfs dfs -test ...', flytectl ... | jq -e '.phase=="SUCCEEDED"'. Review once, before any code. After that, "done" is a non-zero exit code or it isn't done. Adversarial reviewers can be sweet-talked. Failing exit codes cannot.

🌙 Overnight Delegated Delivery

"Hand it off at 6 PM. Wake up to a green PR."

Claude Code alone caps at ~20 min before drift. Superteam runs 20+ hours: every increment gets a fresh Generator/Evaluator pair, and the Evaluator can't read the Generator's reasoning — adversarial by construction. No accumulated context. No fake "looks good." Pair with tmux on a remote VM and the team runs while you sleep.

🛟 Stalls Heal. Pipelines Resurrect.

"Real Isolated Claude Code sessions in tmux. A file-based harness keeps them in sync. Watchdogs bring them back."

Every teammate is a full Claude Code session in its own tmux pane — not an in-process sub-agent, not a multi-agent prompt. Coordination lives on disk (state.json, events.jsonl, frozen contracts, hooks), so the harness self-heals:

• 1200-s watchdog — auto-relaunches the Orchestrator when state.json goes stale
• 5-strike escalation — changes the approach every strike (retry → nudge → fresh pair → split → user); never just retries
• Phase-4 strict re-eval — fresh Evaluator re-runs all final gates; up to 3 FAIL → fix → re-evaluate cycles before escalating to you

Inspired by Anthropic's harness-design writeup.

🧠 Learns Your Stack Once. Forever.

"Stop teaching it about your company every Monday."