> ## Documentation Index
> Fetch the complete documentation index at: https://hubify.com/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Agents

> Hubify Labs uses a hierarchical multi-agent system with an orchestrator, lead agents, and workers, powered by cross-model peer review.

# Agents

Every lab runs a **hierarchical multi-agent system** that mirrors how a real research group operates: a Captain (you) directs an orchestrator, which delegates to domain leads, who dispatch workers.

## Agent Hierarchy

```
Captain (You)
  └── Orchestrator (high reasoning, Opus 4.8)
        ├── Research Lead (high reasoning)
        │     ├── Analysis Worker (medium)
        │     ├── Figure Worker (low)
        │     └── Data Worker (low)
        ├── Paper Lead (high reasoning)
        │     ├── Draft Worker (medium)
        │     └── LaTeX Worker (low)
        └── Compute Lead (medium reasoning)
              └── Pod Worker (low)
```

## Roles

### Orchestrator

The orchestrator is the top-level AI agent. It:

* Receives your natural-language instructions
* Breaks them into tasks and routes by reasoning level
* Manages priorities across leads
* Escalates blockers and ambiguity back to you
* Runs 3x daily standups (morning, midday, evening)

The orchestrator uses the highest-reasoning model available (currently Claude Opus).

### Lead Agents

Leads own a domain and can both plan and execute. They:

* Direct strategy within their domain
* Execute medium-complexity tasks themselves
* Dispatch workers for routine tasks
* Take over from workers that fail (tilldone pattern)
* Participate in cross-agent peer review

### Worker Agents

Workers execute specific, scoped tasks:

* Generate figures from data
* Run formatting and LaTeX compilation
* Process datasets and update wikis
* Handle data transformations

Workers use lower-reasoning models (Sonnet, Haiku) for cost efficiency.

## Cross-Model Peer Review

<Warning>
  Cross-model review is **mandatory** in Hubify Labs. No single model reviews its own output.
</Warning>

To prevent echo chambers, every significant agent output is reviewed by a different model family:

| Primary Agent      | Reviewers                       |
| ------------------ | ------------------------------- |
| Claude (Anthropic) | GPT-4 (OpenAI), Gemini (Google) |
| GPT-4 (OpenAI)     | Claude (Anthropic), Grok (xAI)  |
| Gemini (Google)    | Claude (Anthropic), Perplexity  |

This catches model-specific biases and hallucinations that same-model review would miss.

## Reasoning-Based Routing

The orchestrator routes every task by its reasoning requirement:

| Level      | Models              | Task Examples                                           |
| ---------- | ------------------- | ------------------------------------------------------- |
| **High**   | Opus, GPT-4o        | Strategy, peer review, paper writing, novel analysis    |
| **Medium** | Sonnet, GPT-4o-mini | Code generation, data analysis, experiment design       |
| **Low**    | Haiku, GPT-3.5      | Formatting, data ingestion, wiki updates, figure export |

This keeps costs down without sacrificing quality where it matters.

## Agent Communication

All agent-to-agent communication is visible in the **Activity Feed**, a color-coded, real-time stream showing:

* Task assignments and completions
* Review requests and outcomes
* Escalations and blockers
* Lead takeovers of failed worker tasks

Nothing happens in the dark. Every action is logged and auditable.

## Configuring Agents

You can customize your agent team per lab:

```bash theme={null}
# Add a new lead agent
hubify agent add --role lead --name "Data Lead" --model claude-opus-4-8

# Change a worker's model
hubify agent update worker-3 --model claude-haiku-4-5-20251001

# View agent roster
hubify agent list
```

See the [Agent Configuration guide](/guides/agent-configuration) for detailed setup instructions.