CrewAI

​

What you get

​

Install

pip install "decimalai>=0.27" "crewai>=1.15.27"
export OPENAI_API_KEY="sk-..."

​

Run with no account

import decimalai.crewai
from crewai import LLM, Agent, Crew, Task

agent = Agent(
    role="Assistant",
    goal="Answer the question",
    backstory="A helpful assistant.",
    llm=LLM(model="gpt-5.4-mini"),
    skills=decimalai.crewai.skills(),      # the installed folder
    tools=decimalai.crewai.skill_tools(),  # reads a skill's files
)
question = (
    "What command connects the browser CLI tool "
    "to a browser already running on http://localhost:9222?"
)
task = Task(
    description=question
    + " Load any skill that could apply before you answer.",
    expected_output="The command",
    agent=agent,
)
print(Crew(agents=[agent], tasks=[task]).kickoff())

• skills=decimalai.crewai.skills() is the installed folder. CrewAI lists every skill in it in its own catalog and gives the agent its own load_skill tool, which returns a skill’s full text.
• tools=decimalai.crewai.skill_tools() adds what CrewAI’s loader lacks. load_skill_resource reads a file under a skill’s references/, assets/ or scripts/. run_skill_script runs one of its scripts/, with no shell and a 60-second limit. Both accept only installed skills and paths inside them. run_skill_script doesn’t enforce a skill’s allowed-tools, and skill_tools(scripts=False) leaves it out.
• CrewAI reads the whole folder. Every skill folder in .agents/skills is offered, whether or not .decimal/skills.toml lists it. agent_config() with no key, on other frameworks, uses only the listed ones.

​

Your own tools

from crewai.tools import tool


@tool("Look up an order")
def lookup_order(order_id: str) -> str:
    """Look up the status of an order."""
    return f"Order {order_id} shipped yesterday."


agent = Agent(
    role="Assistant",
    goal="Answer the question",
    backstory="A helpful assistant.",
    llm=LLM(model="gpt-5.4-mini"),
    skills=decimalai.crewai.skills(),
    tools=[lookup_order, *decimalai.crewai.skill_tools()],
)

​

Add a registry skill

decimalai skills add playwright-cli

​

Add a key

export DECIMAL_API_KEY="dai_sk_..."
pip install "openinference-instrumentation-crewai>=1.1.21" openinference-instrumentation-openai
decimalai skills add playwright-cli --to-agent my-agent

import decimalai
from decimalai.crewai import instrument
from crewai import LLM, Agent, Crew, Task

decimalai.init()                                             # reads DECIMAL_API_KEY
instrument(agent_name="my-agent", enable_skill_loader=True)  # before the first kickoff()

agent = Agent(
    role="Assistant",
    goal="Answer the question",
    backstory="A helpful assistant.",
    llm=LLM(model="gpt-5.4-mini"),
)
task = Task(description=question, expected_output="The command", agent=agent)
print(Crew(agents=[agent], tasks=[task]).kickoff())
decimalai.flush()                                            # send the trace now

• enable_skill_loader=True hooks CrewAI’s before_llm_call. There’s no load_skill on this route, so each model call an agent makes is ranked on the task’s description, and the skill list goes into that call right after CrewAI’s system prompt, with the full text of the skill ranked first. If the ranking wasn’t about this task and the agent has other skills, the call gets the list alone, and the model can’t open a skill from it. Calls outside an agent’s turn, such as output conversion and guardrails, are left alone.
• decimalai.init() comes first. Without it, instrument() installs and the agent gets no skills, with no error. decimalai.init(crewai=True, agent_name="my-agent") is the one-line form of both calls.
• The agent has no skills=. An agent with CrewAI skills of its own keeps them, and this route stays out of its calls.
• Call it before the first kickoff(). CrewAI copies its hooks into an agent’s executor when it builds one, so an executor built earlier gets no skills.
• The two packages are for tracing. Skills reach the model without them, but no run is recorded. Use openinference-instrumentation-anthropic or -google-genai for those models.
• instrument() doesn’t connect the project. It copies nothing from .decimal/skills.toml and writes no .decimal/link.toml, which is why the skill goes on with --to-agent or on the agent’s page. Folders in .agents/skills go unused on this route.

​

Change the prompt without redeploying

support = decimalai.agent_config("my-agent", fallback="A helpful assistant.")
agent = Agent(
    role="Assistant",
    goal="Answer the question",
    backstory=support.prompt,
    llm=LLM(model="gpt-5.4-mini"),
)

​

Conversations

​

Its own skill loader

decimalai.init(crewai=True, agent_name="my-agent")

• Each load counts as Loaded, with the version of the text the model received, and the names in CrewAI’s catalog count as Offered.
• Updates take effect at once. CrewAI reads a skill’s text from disk on every load, so after decimalai skills update the same agent serves the new version.
• DecimalAI’s route steps aside by itself. That flag also turns on the instrument() route, which leaves an agent with skills= alone, so there’s no need for skills=False. If another agent in the same crew gets skills that way, a warning says once that two catalogs reached the model in that run.

​

Tracing

• openinference-instrumentation-crewai records the crew, each task, agent interactions and tool calls, with inputs and outputs.
• A provider instrumentor (openinference-instrumentation-openai, -anthropic or -google-genai) records the model calls, with the model name, token counts and messages. CrewAI’s own instrumentor emits no model call.
• Skills on the instrument() route are recorded as Offered and Delivered, never Loaded, because the model never asks for one.
• To trace with no skills, use decimalai.init(crewai=True, skills=False), or set DECIMALAI_SKILLS=0.
• instrument(on_trace=fn) calls fn once per run, failed runs included, with a copy of the finished trace before it’s sent. A run inside decimalai.start_trace() is filed as that trace’s child.
• Several agents in one process. Wrap a kickoff in agent_run("billing-crew"), from decimalai.providers, to file the run and route its skills under that name.

​

Troubleshooting

• The agent ignored its skill. Have the task ask it to load one, as main.py does. See the measurement in Run with no account.
• SkillsNotInstalledError. The project has no skills yet. Run decimalai skills add in its folder.
• A skill is missing from CrewAI’s catalog. CrewAI skips a folder it can’t parse, and logs it on crewai.skills instead of raising. Its metadata parser requires every value to be a string. decimalai skills verify reports what a loader would reject.
• NO CrewAI traces will be captured. openinference-instrumentation-crewai is missing, or refused your CrewAI version. The warning says which.
• The trace has crew and task spans but no model or tokens. The provider instrumentor is missing. It’s only logged at INFO, as CrewAI traces will lack openai model/token detail.
• No skills arrive on the instrument() route, and nothing is logged. Check for decimalai.init() before the first kickoff(), and for skills= on the agent. That one is logged at INFO, once per agent.
• LiteLLM calls aren’t traced. DecimalAI turns on the OpenAI, Anthropic and Google instrumentors only. Install openinference-instrumentation-litellm and call LiteLLMInstrumentor().instrument() yourself.
• CrewAI ran its before_llm_call hook on a thread with a running asyncio event loop. The deprecated CrewAgentExecutor under Crew.akickoff() blocks the loop during ranking. Drop executor_class, or use Crew.kickoff_async().
• enable_load_skill_tool is not supported on the crewai adapter. The argument does nothing, because a hook can’t add a tool. For a model that asks for skills, use its own skill loader.