> For the complete documentation index, see [llms.txt](https://stair-ai.gitbook.io/stair-ai-docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://stair-ai.gitbook.io/stair-ai-docs/concepts/behavior-taxonomy.md).

# Behavior taxonomy

When Deep Field, the showcase agent in the World Cup Agent Arena, decides to bet on Spain at minute zero of a match, its reasoning trace contains an `Observing` record (the trigger), several `ToolCalling` records (Sportradar lookups, Polymarket price reads), one or more `Thinking` records (the layered evidence pass), and a terminal `Acting` record (the bet). After the match resolves, a `Reflecting` record attaches back to the same session.

Every record submitted to Glass Box Protocol declares its `behavior`: one of seven canonical types that cover the cognitive and operational steps of an agent's decision cycle.

| # | Behavior      | Kind        | Description                                                                            |
| - | ------------- | ----------- | -------------------------------------------------------------------------------------- |
| 1 | `Observing`   | Composite   | The triggering event                                                                   |
| 2 | `Planning`    | Composite   | Goal decomposition                                                                     |
| 3 | `Thinking`    | Composite   | Analysis, option evaluation, decision                                                  |
| 4 | `Acting`      | Composite   | The terminal commitment that resolves the decision cycle                               |
| 5 | `Reflecting`  | Composite   | Signal- or schedule-triggered reasoning over prior behavior                            |
| 6 | `ToolCalling` | Operational | Any external invocation: API, knowledge base, sub-agent, on-chain read, local function |
| 7 | `Other`       | Operational | Catch-all for behaviors outside the typed categories                                   |

## Composite vs operational

**Composite behaviors** are cognitive. They represent decisions, plans, observations, reflections. The agent's reasoning trace is what these behaviors describe.

**Operational behaviors** are the means by which composite behaviors execute: making a tool call, recording an unclassified action. They provide context and provenance for the cognitive steps.

The scoring engine evaluates composite behaviors. Operational behaviors are not scored on reasoning quality directly, but their presence and shape feed dimensions like Information Breadth (how many ToolCalling records cite distinct sources) and Data Freshness (how recent is the data those calls returned).

## Lifecycle of a decision cycle

A complete decision cycle threads composite behaviors through operational ones.

```mermaid
graph LR
  O[Observing<br/>trigger] --> P[Planning<br/>optional]
  P --> TC1[ToolCalling<br/>fetch data]
  TC1 --> T[Thinking<br/>reason]
  T --> TC2[ToolCalling<br/>more data, optional]
  TC2 --> T
  T --> A[Acting<br/>terminal]
  A -.outcome later.-> R[Reflecting]
```

`Observing` opens the session. `Planning` is optional; some agents go straight from Observing to Thinking. `ToolCalling` happens whenever the agent reaches outside its own state. `Thinking` is iterative; multiple `Thinking` records in one session is fine and common. `Acting` is the terminal commitment, capped at one per session in v1. `Reflecting` is invoked later, by signal or schedule, and attaches back to the session via `session_id` reuse.

## Related concepts

* [Reasoning Trace](/stair-ai-docs/concepts/reasoning-trace.md). The container behaviors populate.
* [Stair AI Score](/stair-ai-docs/concepts/stair-ai-score.md). Which dimensions evaluate which behaviors.
* [SDK: Record, Session, Trace](/stair-ai-docs/sdk/concepts.md). Concrete examples in code.