| name | autobeat |
| description | Use when delegating work to background agents, creating workflows, choosing between autobeat primitives (tasks, pipelines, loops, orchestrations, schedules), monitoring progress, or troubleshooting failures. |
| user-invocable | false |
| allowed-tools | Read, Grep, Glob |
Autobeat Agent Orchestration
Autobeat lets you delegate work to background AI agent instances, build task pipelines,
create iterative loops, schedule recurring work, and run autonomous orchestrations.
Two runtimes supported: Claude, Codex.
Iron Law
USE THE SIMPLEST PRIMITIVE THAT FITS
Single task beats pipeline. Pipeline beats loop. Loop beats orchestrator.
Each level adds overhead and complexity — only escalate when the simpler
primitive genuinely cannot express your intent. If you catch yourself building
a pipeline with one step, use a task. If you're manually wiring 5 sequential
tasks with dependsOn, use a pipeline.
When This Skill Activates
- Delegating work to background agents
- Choosing between tasks, pipelines, loops, schedules, or orchestrations
- Building multi-step workflows or dependency graphs
- Setting up iterative improvement (retry/optimize loops)
- Monitoring task progress or troubleshooting failures
- Scheduling recurring work
Capability Hierarchy
Use this decision tree to pick the right primitive:
Is it a single, self-contained piece of work?
YES → DelegateTask (single task)
Is it a fixed sequence of steps?
YES → How many steps?
2-20 → CreatePipeline
>20 → Break into multiple pipelines or use orchestrator
Does it need iterative improvement?
YES → Is the exit condition objective (shell exit code / script score)?
YES → CreateLoop with evalMode: shell
NO → CreateLoop with evalMode: agent
Which eval sub-strategy?
Need findings-only feedback, always run to maxIterations? → evalType: feedforward (default)
Need AI judge to decide continue/stop? → evalType: judge
Need deterministic structured pass/fail (Claude only)? → evalType: schema
Is the goal open-ended and complex?
YES → CreateOrchestrator (autonomous planning + delegation)
Should it run on a schedule?
YES → Wrap any of the above: ScheduleTask, SchedulePipeline, ScheduleLoop
Primitive Comparison
| Primitive | Use When | Complexity | Autonomy |
|---|
| Task | Single work item | Lowest | None |
| Pipeline | Fixed sequence (2-20 steps) | Low | None |
| Loop | Iterative improvement | Medium | Exit condition only |
| Orchestrator | Open-ended goals | Highest | Full (plans + delegates) |
| Schedule | Any of the above, recurring/deferred | +1 layer | Timer-driven |
Quick Reference
MCP Tools
| Tool | Purpose |
|---|
DelegateTask | Run a single task in a background agent |
TaskStatus | Check task status (omit taskId for all) |
TaskLogs | Read stdout/stderr from a task |
CancelTask | Cancel a running task |
RetryTask | Re-run a failed/completed task |
ResumeTask | Resume from checkpoint with context |
CreatePipeline | Sequential task chain (2-20 steps) |
CreateLoop | Iterative retry/optimize loop |
LoopStatus | Check loop progress and history |
ListLoops | List loops with status filter |
CancelLoop | Cancel an active loop |
PauseLoop | Pause a loop mid-iteration |
ResumeLoop | Resume a paused loop |
ScheduleTask | Schedule a task (cron or one-time) |
SchedulePipeline | Schedule a recurring or one-time pipeline |
ScheduleLoop | Schedule a recurring loop |
ListSchedules | List schedules with status filter |
ScheduleStatus | Get schedule details + history |
PauseSchedule | Pause a schedule |
ResumeSchedule | Resume a paused schedule |
CancelSchedule | Cancel a schedule |
CreateOrchestrator | Autonomous goal execution |
OrchestratorStatus | Check orchestration progress |
ListOrchestrators | List orchestrations |
CancelOrchestrator | Cancel an orchestration |
ListAgents | List available agents with auth status |
ConfigureAgent | Check auth, store/reset API keys |
InitCustomOrchestrator | Scaffold custom orchestrator (state file, exit script, snippets) |
PipelineStatus | Check pipeline entity status |
ListPipelines | List pipelines with status filter |
CancelPipeline | Cancel a pipeline and optionally its tasks |
CLI Commands
| Command | Purpose |
|---|
beat run "<prompt>" | Delegate a task |
beat status [task-id] | Check task status |
beat logs <task-id> | Read task output |
beat cancel <task-id> | Cancel a task |
beat resume <task-id> | Resume from checkpoint |
beat pipeline "<step1>" --delay 5m "<step2>" | Create pipeline |
beat loop "<prompt>" --until "<cmd>" | Retry loop |
beat loop "<prompt>" --eval "<cmd>" --maximize | Optimize loop |
beat loop "<prompt>" --eval-mode agent --strategy retry | Agent eval loop |
beat schedule create "<prompt>" --cron "0 9 * * *" | Cron schedule |
beat orchestrate "<goal>" | Start orchestration |
beat orchestrate status <id> | Check orchestration |
beat orchestrate init "<goal>" | Scaffold custom orchestrator |
beat dashboard / beat dash | Terminal dashboard TUI |
beat list / beat ls | List tasks |
beat agents check | Check agent auth status |
beat agents config set/show/reset | Agent config management |
beat config show/set/reset/path | Config management |
Composition Patterns
Pipeline-in-Loop: Repeat a multi-step pipeline until quality passes.
CreateLoop { pipelineSteps: ["lint", "test", "build"], strategy: "retry", exitCondition: "npm test" }
Loop-in-Schedule: Run an optimization loop daily.
ScheduleLoop { strategy: "optimize", cronExpression: "0 2 * * *", ... }
Task Dependencies (manual DAG): Fan-out, then fan-in.
A = DelegateTask("generate data")
B = DelegateTask("process subset 1", dependsOn: [A])
C = DelegateTask("process subset 2", dependsOn: [A])
D = DelegateTask("merge results", dependsOn: [B, C], continueFrom: B)
System Prompts
Inject custom instructions into any agent. Identical mechanics across all tools that accept systemPrompt.
| Agent | Mechanism | Behavior |
|---|
| Claude | --append-system-prompt | Appends to Claude's built-in system prompt |
| Codex | -c developer_instructions | Sets developer instructions config |
Supported on: DelegateTask, CreatePipeline (pipeline + per-step), CreateLoop, ScheduleTask, SchedulePipeline (pipeline + per-step), ScheduleLoop, CreateOrchestrator.
Caveat: CreateOrchestrator systemPrompt replaces auto-generated role instructions entirely (not appends). This prevents conflicting ROLE sections. Use InitCustomOrchestrator for custom orchestrators with full prompt control.
Model Selection
Override the default model per task, per pipeline step, or per loop iteration.
Resolution order: per-task model > agent-config default (ConfigureAgent) > agent's built-in default.
Pipelines support per-step overrides: set model at pipeline level for the default, override on individual steps[] as needed.
Model names are opaque to autobeat — validation is the agent CLI's responsibility.
Translation Proxy & Ollama Runtime
Configure via ConfigureAgent (set action):
proxy: "openai": Routes Anthropic API calls through a local proxy that translates to OpenAI-compatible format. Requires baseUrl and apiKey. Works with all agents.
runtime: "ollama": Wraps agent spawns with ollama launch. Supported agents: claude, codex.
Mutually exclusive — runtime takes precedence if both are set. Clear with empty string: proxy: "" or runtime: "".
Anti-Patterns
| Mistake | Why It's Wrong | Fix |
|---|
| Pipeline with 1 step | Unnecessary overhead | Use DelegateTask |
| Manual dependsOn chain for sequential tasks | Error-prone wiring | Use CreatePipeline |
| Loop without exit condition (shell mode) | Runs forever | Set exitCondition, or use evalMode: agent with an evalType |
| Orchestrator for simple sequences | Overkill | Use Pipeline or Loop |
| Polling TaskStatus in a tight loop | Wastes resources | Check periodically (30s+) |
| Ignoring workingDirectory | Tasks run in wrong directory | Always set workingDirectory |
| Unlimited maxIterations with no failures cap | Risk of infinite loop | Set maxIterations or maxConsecutiveFailures |
| Using CreateOrchestrator for custom eval | Auto-generated prompts don't support custom eval logic | Use InitCustomOrchestrator + CreateLoop |
Extended References
Load these for deep dives on specific capabilities:
- orchestration.md — Choosing primitives, composition patterns, orchestrator guardrails. Load when planning complex workflows.
- loops.md — Retry/optimize strategies, agent eval mode, git integration, recipes. Load when creating loops.
- dependencies.md — DAGs, pipelines, context passing, failure cascade. Load when wiring task dependencies.
- monitoring.md — Status checking, recovery, troubleshooting. Load when monitoring or debugging.
- capability-matrix.md — Complete parameter tables for all MCP tools and CLI commands. Load for exact parameter names and defaults.