> For the complete documentation index, see [llms.txt](https://tinyhumans.gitbook.io/openhuman/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://tinyhumans.gitbook.io/openhuman/features/orchestration.md).

# The Orchestrator

OpenHuman is an orchestrator, not a chatbot: durable agent graphs, visual workflows, sub-agent fleets, and a split-brain always-on layer, all in one coherent stack.

<figure><img src="https://1136424271-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2Fe42lJPkxRsAFpQOln2KW%2Fuploads%2Fgit-blob-93a9131dd351026170663db5c4447e8eb162746e%2Forchestration.png?alt=media" alt=""><figcaption><p>OpenHuman orchestrating a fleet of agents.</p></figcaption></figure>

Most harnesses run one agent in one loop. OpenHuman is built as an **orchestrator**: a stack for coordinating many agents, over long horizons, across machines. It does this durably, observably, and under your control.

Four layers make that real:

## 1. Graphs, not loops

Every agent turn runs on [tinyagents](https://github.com/tinyhumansai/tinyagents), our open-source graph engine. Multi-step work compiles to **state-machine graphs with conditional routing**: `plan → execute ⇄ review → finalize` for delegation, phase DAGs for multi-agent workflow runs, and map-reduce fan-out for parallel workers. All of it has **durable checkpointing**. A graph can pause mid-run (for your answer, for an approval, for a restart) and resume exactly where it stopped.

## 2. Sub-agent fleets that don't get lost

The orchestrator spawns specialized sub-agents (up to 3 levels deep), reuses compatible idle workers instead of re-spawning, and routes each to the right model tier: heavy reasoning for the core, a fast **burst tier** for low-context workers. Reliability is structural: a no-progress circuit breaker stops loops, and stuck children hand back a `question` (pause + resume on your answer) or an `Incomplete` root-cause summary, never silence. See the [Agent Harness](/openhuman/developing/architecture/agent-harness.md).

## 3. Workflows you can see

[Workflows](/openhuman/features/workflows.md) lift orchestration out of the chat: the agent *proposes* a typed graph of triggers, agents, tools and conditions; you review it on a canvas and save it. Runs are durable, approval-gated, and fully inspectable step-by-step, powered by open-source [tinyflows](https://github.com/tinyhumansai/tinyflows).

## 4. An always-on split brain

Inbound traffic hits a **fast reflex agent** that triages in seconds and hands a deep **reasoning core** a concise brief; the core does the multi-step work and delegates to workers. Mid-task steering input can be delivered into a live session, and 20:1 compression keeps week-long sessions bounded.

## What's next

The scripted-orchestration idea this page used to advertise, agents writing control flow as small programs in a sandboxed REPL, is **not being built**: the harness dropped that runtime and the tool no longer exists. The substrate above, graphs with checkpointing under a trust model, is the shape orchestration takes.

What is tracked in the open instead: an [agent-to-agent protocol](https://github.com/tinyhumansai/openhuman/issues/3463) so a graph can span instances, and durable swarm task graphs. See the [roadmap](/openhuman/overview/roadmap.md).

***

## Why this differentiates

|                 | Single-agent harnesses (Claude Code, OpenClaw, Hermes) | OpenHuman                                                 |
| --------------- | ------------------------------------------------------ | --------------------------------------------------------- |
| Execution model | One loop, one context                                  | Compiled graphs, conditional routing, checkpoint/resume   |
| Parallelism     | Manual / plugin                                        | Native sub-agent fleets, map-reduce fan-out, worker reuse |
| Automation      | Scripts & cron                                         | Visual, durable, approval-gated workflows                 |
| Always-on       | None                                                   | Split-brain reflex + reasoning core, mid-task steering    |

## See also

* [Workflows](/openhuman/features/workflows.md)
* [Agent Harness](/openhuman/developing/architecture/agent-harness.md): the developer deep-dive on graphs, breakers, journals.
* [Agent Coordination tools](/openhuman/features/native-tools/agent-coordination.md): the user-facing spawn/delegate surface.