agent-orchestration
npx machina-cli add skill parcadei/Continuous-Claude-v3/agent-orchestration --openclawAgent Orchestration Rules
When the user asks to implement something, use implementation agents to preserve main context.
The Pattern
Wrong - burns context:
Main: Read files → Understand → Make edits → Report
(2000+ tokens consumed in main context)
Right - preserves context:
Main: Spawn agent("implement X per plan")
↓
Agent: Reads files → Understands → Edits → Tests
↓
Main: Gets summary (~200 tokens)
When to Use Agents
| Task Type | Use Agent? | Reason |
|---|---|---|
| Multi-file implementation | Yes | Agent handles complexity internally |
| Following a plan phase | Yes | Agent reads plan, implements |
| New feature with tests | Yes | Agent can run tests |
| Single-line fix | No | Faster to do directly |
| Quick config change | No | Overhead not worth it |
Key Insight
Agents read their own context. Don't read files in main chat just to understand what to pass to an agent - give them the task and they figure it out.
Example Prompt
Implement Phase 4: Outcome Marking Hook from the Artifact Index plan.
**Plan location:** thoughts/shared/plans/2025-12-24-artifact-index.md (search for "Phase 4")
**What to create:**
1. TypeScript hook
2. Shell wrapper
3. Python script
4. Register in settings.json
When done, provide a summary of files created and any issues.
Trigger Words
When user says these, consider using an agent:
- "implement", "build", "create feature"
- "follow the plan", "do phase X"
- "use implementation agents"
Source
git clone https://github.com/parcadei/Continuous-Claude-v3/blob/main/.claude/skills/agent-orchestration/SKILL.mdView on GitHub Overview
Agent Orchestration Rules define how to use implementation agents to preserve the main context when a user asks for an implementation. It emphasizes spawning an agent to perform reading, understanding, editing, and testing, then returning a concise summary to the main context. This helps manage complexity and maintain context across multi-file changes.
How This Skill Works
When a task is requested, the system spawns an agent (e.g., implement X per plan). The agent reads relevant files, understands the requirements, makes edits, and runs tests. After completion, the main context receives a summary (about ~200 tokens) of what was done and any issues encountered.
When to Use It
- Multi-file implementation: use an agent to manage complexity across files.
- Following a plan phase: have the agent read and implement according to a plan.
- New feature with tests: agents can implement features and run tests.
- Single-line fix: avoid agent overhead for speed; do it directly.
- Quick config change: overhead of agents may not be worth it for small tweaks.
Quick Start
- Step 1: Spawn an implementation agent with the task (e.g., agent("implement X per plan")).
- Step 2: Let the agent read files, understand requirements, edit code, and run tests.
- Step 3: Retrieve the main summary (~200 tokens) detailing what was created and any issues.
Best Practices
- Use agents for multi-file implementations to manage internal complexity.
- Always have agents read and follow an explicit plan phase rather than passing raw instructions in the main chat.
- Leverage agents for features that require tests, so edits and test runs are handled automatically.
- For quick, simple changes (single-line fixes or config tweaks), perform the task directly to avoid overhead.
- Require and review a concise summary of results (and any issues) after the agent completes.
Example Use Cases
- Example: Implement Phase 4: Outcome Marking Hook from the Artifact Index plan. Plan location: thoughts/shared/plans/2025-12-24-artifact-index.md (search for 'Phase 4'). What to create: 1) TypeScript hook 2) Shell wrapper 3) Python script 4) Register in settings.json. When done, provide a summary of files created and any issues.
- Example: Spawn agent("implement X per plan"); Agent reads files → understands → edits → tests; Main receives a ~200-token summary.
- Example: Following a plan phase: Agent reads the plan, implements according to said plan, and returns progress and any blockers to the main context.
- Example: New feature with tests: Agent creates code, adds tests, runs test suite, and reports back with test results and coverage notes.
- Example: Quick config change: For a tiny config tweak, skip the orchestration overhead and apply directly in the main chat to save time.
