wisbech

Bandit — AI skill for Claude Code

AI community

The agent factory — AI coding agents on a kanban board, an adversarial critic enforcing quality, verification-gated everything.

How to install Bandit

This entry records only its repository, not the path inside it, so there is no exact command to give. Open wisbech/bandit and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Bandit does

The agent factory — AI coding agents on a kanban board, an adversarial critic enforcing quality, verification-gated everything. Works with opencode, Claude Code, Codex, pi; ACP adapters; any ollama model.

Alternatives in AI

Fast Agent — Code, Build and Evaluate agents - excellent Model and Skills/MCP/ACP Support 3.7k ★
Agent Teams AI — You're the boss, agents are your team 2k ★
Pilotfish — Multi-model orchestration layer for Claude Code — the frontier model plans, cheaper models execute, verificati 688 ★

README

bandit

**The agent factory.** A dark factory where AI coding agents execute work on a kanban board, an adversarial critic enforces quality, and every claim is backed by verification evidence — not vibes.

   ┌─────────────────────────────────────────────────┐
   │                    THE LOOP                     │
   │                                                 │
   │   board ──▶ master ──▶ actor ──▶ verify gate    │
   │    ▲                                   │        │
   │    │                                   ▼        │
   │    └────────────────────────── critic ◀────────┘
   │         verdict: pass / fail / bypass           │
   └─────────────────────────────────────────────────┘

**A serf is a folder. A card is a folder. Events are the truth. Harnesses are adapters.**

bandit coordinates multiple AI coding harnesses (opencode, Claude Code, Codex, pi…) as workers on a shared board, measures whether their work moves the needle, and governs its own confidence about what works. It is the third generation of the serf idea — proven across years of daily production use.

Why

Agent harnesses are powerful but unmanaged. You babysit one session at a time, paste context by hand, and trust output that was never verified. bandit flips that:

Cards instead of conversations. Work is a kanban card with acceptance criteria, not a chat transcript.
Verification instead of vibes. Every actor must report a VERIFICATION_COMMAND and its exit code. Green or it didn't happen.
A critic with teeth. A second adversarial agent evaluates every result against the card. Plumbing failures retry the critic, never the actor.
Events are the truth. Append-only JSONL; the board is a projection. Replay repairs a corrupted board — no state to corrupt that isn't reconstructible.
Bounded autonomy. Convergence rounds (default 3), budget ha