Execute pending agent-surface tasks by dispatching dedicated agents for each task. Quality-first approach — each task gets a focused agent, batched 3-4 in parallel. Use when user says "execute agent tasks", "run agent queue", "process agent-surface tasks", "agent task execution", or during the agent-task-executor cron cycle.
Standardmäßig ist der Prompt ausgewählt, der zuerst die Quelle prüft. Sie können zu einem direkten Befehl wechseln oder eine lokale Kopie herunterladen.
Quelldateien prüfen
Lesen Sie SKILL.md und alle von SkillsMP angezeigten Begleitdateien, bevor Sie sich für eine Installation entscheiden.
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Ein direkter Befehl überspringt den Prüf-Prompt. Prüfen Sie die Quelle, bevor Sie ihn ausführen.
Execute pending agent-surface tasks by dispatching dedicated agents for each task. Quality-first approach — each task gets a focused agent, batched 3-4 in parallel. Use when user says "execute agent tasks", "run agent queue", "process agent-surface tasks", "agent task execution", or during the agent-task-executor cron cycle.
Agent-Task Executor Skill
Execute pending surface: "agent" tasks by dispatching dedicated agents for each one. Every task gets a focused, fresh agent — no bulk processing, no quality loss.
Core Principles
Quality over speed — each task gets a dedicated agent with full focus and context
Parallel batches of 3-4 — launch a batch, wait for completion, then next batch
Ownership via created_by — the agent that created the task is the agent type that executes it
No bulk execution — never try to run all pending tasks at once
Skip what you can't handle — if a task requires human input or has a future due date, leave it
Execution Protocol
Step 1: Query Pending Agent-Surface Tasks
list_tasks(surface="agent", status="pending")
Step 2: Filter & Prioritize
From the pending tasks, build an execution queue by applying these filters:
SKIP these tasks (do NOT execute):
created_by: "user" — these were created by a parent/user and need their timing/direction
Tasks with due_date in the future (not yet actionable)
Tasks that require human judgment (purchase decisions, medical, content approval)
Tasks that reference tools/access you don't have (e.g., browser re-auth)
PRIORITIZE by:
priority: "urgent" first
priority: "high" second
Overdue tasks (past due_date) third
priority: "medium" fourth
priority: "low" last
Step 3: Select a Batch
Pick the top 3-4 tasks from the filtered/prioritized queue. This is one execution batch.
Batch size rules:
Default: 3 tasks per batch
If tasks are simple/quick (config fixes, memory trims): 4 per batch
If tasks are complex (skill extraction, refactoring): 2-3 per batch
Never exceed 4 per batch
If fewer than 3 tasks are available, just run what's there
Step 4: Dispatch Dedicated Agents
For each task in the batch, launch a dedicated agent via the task tool:
All agents in the batch are launched IN PARALLEL. They run simultaneously.
Agent Type Resolution
Map created_by to the correct task tool agent_type:
created_by
agent_type to use
platform-manager
platform-manager
context-auditor
context-auditor
skill-optimizer
skill-optimizer
content-scheduler
content-scheduler
content-manager
content-manager
content-analytics
content-analytics
content-creative
content-creative
coding-agent
coding-agent
health-coach
health-coach
nicu-care
nicu-care
finance-manager
finance-manager
home-manager
home-manager
teacher
teacher
realtor-team
realtor-team
wellness-coach
wellness-coach
parenting-coach
parenting-coach
checkin
general-purpose
heartbeat
general-purpose
legacy
general-purpose
(any other)
general-purpose
If the created_by value matches a custom agent type in the task tool's agent list, use it directly. Fall back to general-purpose only for unknown agent types.
Prompt Construction
Each dispatched agent receives a complete, self-contained prompt:
You have ONE task to execute. Focus entirely on this.
TASK: {task.title}
TASK ID: {task.id}
PRIORITY: {task.priority}
CATEGORY: {task.category}
NOTES: {task.notes}
ASSIGNEE: {task.assignee}
INSTRUCTIONS:
1. Load your memory tiers if you have them (data/agents/{agent-name}/core.md, working.md)
2. Execute this task completely — no stubs, no TODOs, no partial work
3. When finished, end your response with a clear status line:
- SUCCESS: "✅ TASK COMPLETE — {brief description of what was done}"
- BLOCKED: "🚫 TASK BLOCKED — {reason why it cannot be completed}"
4. Do NOT call `complete_task` or `update_task` — the parent orchestrator handles task lifecycle
IMPORTANT:
- This task has surface="agent" — it was created FOR an agent to handle autonomously
- Do NOT send Telegram messages to {{PARENT_1}} about this unless the task explicitly requires notification
- Do NOT create new tasks unless the original task's scope genuinely requires subtasks
- Do NOT call `complete_task`, `update_task`, `list_tasks`, or any task management tool — these are NOT available in sub-agent context (SDK limitation). Just do the work and report status.
- Focus. Execute. Report.
Step 5: Wait for Completion & Handle Task Lifecycle
After launching all agents in the batch:
Use read_agent(agent_id, wait=true, timeout=180) for each agent
Parse the agent's final status line:
If "✅ TASK COMPLETE" → call complete_task(id="{task_id}") in the PARENT session
If "🚫 TASK BLOCKED" → call update_task(id="{task_id}", status="blocked", notes="{reason from agent}") in the PARENT session
If agent errored/crashed → call update_task(id="{task_id}", status="blocked", notes="Agent execution failed") in the PARENT session
If an agent times out (180s), check status with read_agent(agent_id, wait=false)
If still running, let it continue — don't kill it
Move on to the next batch
Come back to mark the task done/blocked in the next cycle
CRITICAL: Task lifecycle (complete_task, update_task) is ALWAYS handled by the parent orchestrator (platform-manager), NEVER by the sub-agent. Sub-agents do not have access to task management tools (SDK v1.0.47 limitation).
Step 6: Report Results
After the batch completes, compile a brief summary:
Agent-Task Execution Results:
✅ Completed: {count} — {list of task titles}
🚫 Blocked: {count} — {list with reasons}
⏳ Still running: {count}
📋 Remaining in queue: {remaining_count}
Notification rules:
If ANY tasks were completed → send summary to {{PARENT_1}} via Telegram ({{TELEGRAM_PARENT_1}}) with speak parameter
If ALL tasks were blocked/failed → send brief alert to {{PARENT_1}}
If nothing was actionable (all skipped) → do NOT message {{PARENT_1}}
Step 7: Next Batch (if applicable)
If more tasks remain in the queue AND the cron cycle has capacity:
Max 2 batches per cron cycle (6-8 tasks total maximum)
Wait for batch 1 to complete before starting batch 2
If batch 1 had failures/blocks, still proceed with batch 2 (different tasks)
Platform-Manager Direct Execution
Some tasks are faster to execute directly rather than dispatching an agent. The platform-manager (or whatever agent invokes this skill) MAY execute these task types directly:
Task Type
Direct Execution OK?
Fix stale file references
✅ Yes — quick file edit
Update cron.json
✅ Yes — config change
Trim/prune memory files
✅ Yes — file edit
Fix JSON syntax
✅ Yes — quick fix
Create calendar events
✅ Yes — single tool call
Skill extraction
❌ No — dedicated agent
Agent refactoring
❌ No — dedicated agent
Multi-file changes
❌ No — dedicated agent
Content review/scheduling
❌ No — dedicated agent
If platform-manager handles a task directly:
Execute the fix
Call complete_task(id="{task_id}")
Include it in the results summary
Anti-Patterns (NEVER do these)
❌ Batch-executing 10+ tasks in a single agent prompt ("handle all these...")
❌ Executing user-created tasks without {{PARENT_1}}'s explicit go-ahead
❌ Expecting sub-agents to call complete_task — they can't (SDK limitation). Parent handles it.
❌ Creating new agent-surface tasks as "execution notes" — just do the work
❌ Sending a Telegram for every individual task completion — batch the report
❌ Executing tasks with future due dates that aren't yet actionable
❌ Running more than 2 batches (8 tasks) per cron cycle — quality degrades with volume
❌ Leaving tasks in "pending" after attempting them — either "done" or "blocked"
Cron Integration
This skill is invoked by the agent-task-executor cron job in cron.json:
{"id":"agent-task-executor","schedule":"0 7,9,11,13,15,17,19 * * *","enabled":true,"agent":"platform-manager","prompt":"Invoke the agent-task-executor skill. Execute pending agent-surface tasks following the skill protocol: query → filter → prioritize → batch (3-4) → dispatch dedicated agents → wait → report. Max 2 batches this cycle."}
Edge Cases
Task references a deleted/renamed agent
If created_by maps to an agent that no longer exists (e.g., a decommissioned team agent), use general-purpose and include extra context in the prompt about what the task is trying to accomplish.
Task is self-referential (created_by = "platform-manager" and executor IS platform-manager)
Execute directly if it's a simple fix. If complex, still dispatch a fresh platform-manager agent — clean context is always better than trying to do everything inline.
Task has dependencies (depends_on other tasks)
Check get_dependency_tree(task_id). If upstream dependencies are not done, skip the task this cycle. It will be picked up when dependencies clear.
Task has been pending for 14+ days
Flag as potentially stale. If the task references a specific date/event that has passed, cancel it with a note. If it's evergreen work, keep it and let the dedicated agent decide.
