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

# Track CrewAI agent tasks as Precipiq decisions

> Pass PrecipiqCrewAICallback as a CrewAI step_callback to record every AgentAction and AgentFinish as a tamper-evident Precipiq decision.

The `precipiq.integrations.crewai` module wraps CrewAI's `step_callback` hook so every `AgentAction` (a mid-task tool call) and `AgentFinish` (task complete) is automatically recorded as a Precipiq decision. Pass the callback when you construct a `crewai.Agent` and all subsequent task executions land in your ledger with full input/output capture and a reasoning trace.

Like the LangChain adapter, this module is **import-time optional** — the base `precipiq` package never imports anything from `crewai`.

## Install

<CodeGroup>
  ```bash pip theme={null}
  pip install precipiq crewai
  ```
</CodeGroup>

## Usage

<CodeGroup>
  ```python Single agent theme={null}
  from crewai import Agent, Crew, Task
  from precipiq import Precipiq
  from precipiq.integrations.crewai import PrecipiqCrewAICallback

  pq = Precipiq(api_key="pq_test_demo_key_REPLACE_ME")

  callback = PrecipiqCrewAICallback(pq, agent_id="researcher")

  researcher = Agent(
      role="Researcher",
      goal="Find three sources on a topic",
      backstory="…",
      step_callback=callback,
  )

  task = Task(
      description="Research the AI Consequences Ledger",
      agent=researcher,
      expected_output="three bullet points",
  )

  Crew(agents=[researcher], tasks=[task]).kickoff()
  ```
</CodeGroup>

## What gets recorded per step

Every `step_callback` invocation produces a decision record with the following fields:

* **`action_type`** — `"agent_action"` for an `AgentAction` (mid-task tool call), or `"agent_finish"` for an `AgentFinish` (task complete).
* **`inputs`** — the `tool_input` on an action, or the prior output on a finish.
* **`outputs`** — the tool's result on an action, or the agent's final answer on a finish.
* **`metadata`** — the CrewAI step `log`, which contains the agent's reasoning trace.

## Multi-agent crews

Instantiate one callback **per agent** so each agent's decisions are attributed to its own `agent_id`. All decisions land on the same org-wide ledger — use the dashboard's agent filter to separate them for reporting.

<CodeGroup>
  ```python Multi-agent crew theme={null}
  researcher_cb = PrecipiqCrewAICallback(pq, agent_id="researcher")
  writer_cb = PrecipiqCrewAICallback(pq, agent_id="writer")

  researcher = Agent(
      role="Researcher",
      step_callback=researcher_cb,
      goal="Find three primary sources on the topic",
      backstory="A careful reader.",
  )
  writer = Agent(
      role="Writer",
      step_callback=writer_cb,
      goal="Turn research into a one-page brief",
      backstory="A clear prose stylist.",
  )
  ```
</CodeGroup>

<Tip>
  With multiple agents in the ledger, the Precipiq dashboard's **agent filter** lets you drill down to the decisions made by a single agent while still viewing the aggregate P\&L across the full crew.
</Tip>