Tasks
Describe what you need in plain English and let agents handle the rest
What It Is
A Task is a complex piece of work that OverClaw breaks down, plans, and executes autonomously through a 6-phase pipeline. Unlike chat (back-and-forth conversation), Tasks run independently — you describe the goal, agents figure out how to get there, and deliver complete results.
How to Use It
Creating a Task
- Tap "New Task" in your workspace
- Describe what you need (5,000 character limit) — or use voice input with waveform visualization and server-side transcription
- Optionally assign specific agents (or let the Orchestrator auto-delegate)
- Choose from 290+ prebuilt templates via the template picker, or write from scratch
- AVA asks clarifying questions in the Refinement step
- Review and tap "Run"
Scheduling
Tasks support recurring execution via ScheduleMode:
| Mode | Description |
|---|---|
| Once | Run immediately |
| Daily | Run every day at specified time |
| Weekly | Run every week on specified day |
| Monthly | Run every month on specified date |
| Interval | Run every N hours (1h, 4h, 12h) |
| Cron | Custom cron expression |
The 6-Phase Pipeline
| Phase | What Happens | Model |
|---|---|---|
| 1. Breakdown | Analyzes your description, generates a structured JSON graph of subtasks with dependencies | Claude Opus 4.6 |
| 2. Planning | Topological sort (Kahn's algorithm) for dependency order. Maps suggested personalities to actual agents via exact match → keyword match → alias expansion → fallback chain | Internal |
| 3. Precheck | Verifies required OAuth integrations are connected. Blocks task if critical connections missing, prompts user to authorize | Gemini Flash |
| 4. Execution | Executes each subtask in dependency order. Agents call tools as needed with up to 10 tool-call iterations per subtask | Per-task routing |
| 5. Review | Reviews subtask outputs for quality. Can request up to 2 retries with "redo" verdict | Gemini Flash |
| 6. Assembly | Extracts artifacts, links, and connection actions. Writes user-friendly summary and compiles work breakdown showing each subtask, assigned agent, tools used | Claude Opus 4.6 |
Tasks run asynchronously on ECS Fargate. The TaskDetailView polls the backend periodically — 5 seconds for active tasks, 15 seconds for idle, with exponential backoff.
Personality Matching (Phase 2)
The Planning phase maps abstract personality suggestions to actual agents in your workspace:
- Exact match — "researcher" → Researcher agent
- Keyword in job description — searches agent descriptions
- Alias expansion — "developer" → Programmer, "writer" → Copywriter
- First non-orchestrator — any available specialist
- Orchestrator — last resort fallback
Task Monitoring
The TaskDetailView provides real-time visibility:
- Live subtask checklist via SubtaskGraphView
- Paginated activity feed (5 per page) with agent attribution
- "Needs clarification" inline input for agent questions
- Timeout alerts when subtasks exceed time limits
- Stop / Resume / Retry actions with confirmation dialogs
What You Need to Know
- Tasks consume more credits than chat. A complex task with 10+ subtasks involves Claude Opus for breakdown and assembly, Gemini Flash for precheck and review, plus multiple execution calls.
- The hard cap is checked before execution. The Spend Guard middleware enforces hourly ($1/hr) and monthly limits before any expensive operation.
- You can stop a running task. Partial results are saved and visible in the activity feed.
- Files are supported. Upload documents, images, audio, and video. Files stored in S3 with CloudFront CDN.
- Task history retained for 90 days. Activity logs, subtask graphs, and generated artifacts all accessible.
Tips
- Be specific. "Research competitor pricing for AI agent platforms and create a comparison PDF" works better than "look into competitors."
- Use templates. The 290+ prebuilt templates across 9 categories provide proven starting points.
- Review the plan. The subtask breakdown shows exactly what agents will do before execution starts.
- Schedule recurring work. Daily reports, weekly summaries, and monthly audits run automatically.
FAQ
How long do tasks take? Simple tasks (2-3 subtasks): 1-2 minutes. Complex tasks (10+ subtasks): 5-15 minutes. Tasks with web browsing or code execution take longer.
What if a subtask needs my input? The task enters "needs clarification" state. You'll see an inline input field in TaskDetailView to respond. The task resumes after you reply.
Can I reuse a completed task? You can create new tasks from templates or copy the description from a previous task. Task results and artifacts are accessible from the run history.