Build a full Claude Code agent system

bernstein

Schedules your fleet as a DAG in plain Python, zero model in the coordination loop. Every node hits the real Claude CLI in its own git worktree, behind a lint/type/test gate, and merges only if it's green. Trap: heavy platform, and the DAG is hand-authored, not inferred for you.

726 ★ · Apache-2.0

agent-worktree

A git worktree per agent, so five agents in one repo stop trampling each other's files. Dry-run merges first and rolls back on conflict. Trap: it's a primitive, not a scheduler. It doesn't decide how many agents to spawn.

267 ★ · MIT

wshobson/agents

Gives each node a role. A real Claude Code plugin marketplace: 203 specialist subagents across 94 plugins, tiered Opus for architecture, Haiku for fast ops. Trap: install only the roles you'll use, or you burn the context you just saved.

38,185 ★ · MIT

insane-research

The whole pattern, already shipped as one plugin. A 7-phase research graph that fans out sub-agents, then lets a deterministic code gate score every claim. Trap: skip the validate step and synthesis has nothing allowed to cite.

108 ★ · MIT

beads

Memory that survives resets. Replaces the markdown TODO with a real dependency graph over versioned SQL. It returns only unblocked tasks, and insights survive sessions. Trap: don't put its folder on iCloud or Dropbox, cloud sync corrupts the database.

25,603 ★ · MIT

waku-agent

The loop core, small enough to read. The whole agent loop is one file, about 95 lines. No done flag: it ends when the model stops asking for tools. Trap: it swallows tool errors into normal output, so a wrong tool looks correct. That's why the gate exists.

440 ★ · MIT

serena

Context, done right. Claude Code's built-in tool descriptions eat about 16,000 un-editable tokens and bias it toward reading whole files. serena forces symbol-level retrieval: the one function, not the 2,000-line file. Trap: it verifies nothing, it only hands the model better tools.

26,813 ★ · MIT

superpowers

Skills plus discipline. A full methodology, not a pack. Its TDD skill enforces one rule: no production code without a failing test first. Trap: it's persuasion, not a syscall. A model can fake a test and claim it saw red. Real enforcement is the gate.

260,116 ★ · MIT

claude-review-loop

The gate. A Stop hook that blocks the agent from quitting until a second model reviews the work, up to four reviewers in parallel. It fails open by design. Trap: no license in the repo, and it only checks a review file exists, the agent can still skip findings.

706 ★ · no license

workshop

Proof. Captures a real run, re-drives that exact trace against your edited code, and diffs the tool calls until the spans go green. Trap: replay is only safe if you extract a clean entrypoint, or it can hit your real database and billing.

937 ★ · MIT

Can your system take “done” back?

bernstein won't merge a node that fails its gate. beads flips a finished task back to not-ready. The review hook un-finishes a finished session. workshop fails a green trace. A system that can only promote is a burndown chart with extra steps. Draw the graph, but build it out of loops first.