Execute implementation plans in parallel using agent teams with quality gates. Spawns implementer teammates in split panes, orchestrated by a team lead who enforces spec compliance and code quality reviews.
From casaflownpx claudepluginhub casaperks/casaflow --plugin casaflowThis skill uses the workspace's default tool permissions.
implementer-prompt.mdlead-playbook.mdspec-reviewer-prompt.mdDesigns and optimizes AI agent action spaces, tool definitions, observation formats, error recovery, and context for higher task completion rates.
Enables AI agents to execute x402 payments with per-task budgets, spending controls, and non-custodial wallets via MCP tools. Use when agents pay for APIs, services, or other agents.
Compares coding agents like Claude Code and Aider on custom YAML-defined codebase tasks using git worktrees, measuring pass rate, cost, time, and consistency.
Execute implementation plans by spawning implementer teammates that work in parallel, each in their own split pane. The team lead orchestrates a staggered review pipeline — as each implementer finishes, the lead dispatches spec compliance and code quality reviewer subagents. Quality gates are enforced: no task is complete until both reviews pass.
Core principle: Parallel implementation via agent teams + two-stage review gates = fast iteration with high quality.
flowchart TD
A{Have implementation<br/>plan?} -->|No| B[Use brainstorm +<br/>plan first]
A -->|Yes| C{3+ independent<br/>tasks?}
C -->|No| D[Use sdd<br/>sequential]
C -->|Yes| E{Tasks touch mostly<br/>different files?}
E -->|No| F[Use sdd<br/>sequential]
E -->|Yes| G{Agent teams<br/>enabled?}
G -->|No| H[Enable agent teams<br/>or use sdd]
G -->|Yes| I[team-dev]
style I fill:#e8f5e9
| sdd | Parallel Agents | team-dev | |
|---|---|---|---|
| Execution | Sequential, one task at a time | Parallel, fire-and-forget | Parallel, staggered pipeline |
| Quality Gates | Spec + quality review | None | Spec + quality review |
| Visibility | Single session | Background subagents | Split panes, interactive |
| Feedback Loop | Synchronous within session | None | Bidirectional via SendMessage |
| Best For | Sequential or coupled tasks | Quick parallel research | Independent tasks with quality needs |
| Token Cost | Lowest | Medium | Highest (persistent sessions) |
CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS: "1" in settingsteammateMode: "tmux" in ~/.claude/settings.json (recommended for split panes)tmux -CC in iTerm2 or tmux new -s claude then claudeplan)1. Enable agent teams — add to your project's .claude/settings.json:
{
"env": {
"CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
}
}
2. Enable tmux teammate mode — add to your user settings at ~/.claude/settings.json:
{
"teammateMode": "tmux"
}
3. Launch Claude inside tmux:
# iTerm2 (recommended — gives you native iTerm2 tabs/panes)
tmux -CC
# Standard terminal
tmux new -s claude
Then run claude from inside the tmux session. Teammates appear as split panes automatically.
4. Have an implementation plan — use plan to generate one, or write one manually.
flowchart TD
read[1. Read plan<br/>Analyze dependencies + file overlap] --> count[2. Determine implementer count<br/>Bias toward 3]
count --> create[3. TeamCreate + TaskCreate<br/>Set blockedBy relationships]
create --> spawn[4. Spawn implementer teammates<br/>One per initial parallel task]
spawn --> work[Implementer works<br/>Implement, test, self-review]
work --> done[Implementer reports done]
done --> spec[Lead dispatches<br/>spec reviewer subagent]
spec --> spec_pass{Spec passes?}
spec_pass -->|No| spec_fix[Send feedback<br/>Implementer fixes] -.-> spec
spec_pass -->|Yes| quality[Lead dispatches<br/>swarm review - fast-pass]
quality --> qual_pass{Quality passes?}
qual_pass -->|No| qual_fix[Send feedback<br/>Implementer fixes] -.-> quality
qual_pass -->|Yes| complete[Task complete]
complete --> more{More unblocked tasks?}
more -->|Yes| related{Related to<br/>implementer's work?}
more -->|No| shutdown[Shut down implementer]
related -->|Yes| reuse[Reuse implementer<br/>Send next task] --> work
related -->|No| fresh[Shut down + spawn fresh] --> work
shutdown -.->|when all done| all_done[All tasks complete]
all_done --> integration[Final integration review]
integration --> cleanup[Shutdown teammates<br/>TeamDelete]
cleanup --> finish[finish]
| Template | Purpose | Invocation |
|---|---|---|
./implementer-prompt.md | Spawn an implementer teammate | Task tool with team_name |
./spec-reviewer-prompt.md | Dispatch spec compliance reviewer | Subagent (not teammate) |
review skill | Dispatch swarm quality review (fast-pass tier) | Subagent via skill pipeline |
./lead-playbook.md | Lead's orchestration decision logic | Reference for the lead |
Reviews happen as implementers finish — no waiting for all to complete. Multiple reviews can run in parallel since they are read-only.
An implementer does NOT start their next task until both reviews pass on their current task. This prevents messy context switches when review feedback arrives.
Reviewers are subagents (not teammates). They report to the lead, who decides next steps. This keeps the lead as the single quality gatekeeper.
| Independent Task Count | Recommended Implementers |
|---|---|
| 2 | 2 |
| 3-4 | 3 |
| 5-6 | 4-5 |
| 7+ | 5-6 (cap at 6) |
Bias toward 3. Scale up when tasks are truly independent. Scale down when overlap is high.
The lead groups tasks by file surface area before spawning. Tasks touching the same files never run in parallel — they use blockedBy dependencies instead.
Never:
If implementer asks questions:
If reviewer finds issues:
If implementer gets stuck:
Required workflow (in order):
Used during orchestration:
Implementer teammates use:
Alternative workflows:
| Situation | Use Instead |
|---|---|
| Fewer than 3 independent tasks | sdd |
| Tasks share many files | sdd |
| Agent teams not enabled | sdd |