crewai adapter ​
Status: shipped, release phase. It passes the contract suite in chat mode, through its process interface, on a fake bridge that plays the Acme example agent. Its bridge ran end to end on the real crewai 1.15.27 with a scripted model, tools reported from the event bus. The HTTP interface is tested on a fake AMP API built from its documented shape; it has not run against a deployed crew yet.
The crewai adapter runs a CrewAI crew. A crew runs tasks; a conversation reaches it through its inputs. It has two interfaces:
crewaiProcess: the crew in a Python process, through a bridge script that ships with the adapter. It reports the final output, each task's output, and every tool the crew's agents used.crewaiHttp: a crew deployed on CrewAI AMP, throughPOST /kickoffandGET /status/{kickoff_id}. That API reports outputs only, no tool calls.
Python and crewai are yours to install; driftbench depends on neither.
Options ​
import { crewaiAdapter, crewaiHttp, crewaiProcess } from '@agent-nexus/driftbench-adapters';
const local = crewaiAdapter({
transport: crewaiProcess({ crew: 'acme.crews:build_sales', python: '.venv/bin/python', cwd: '/srv/acme' }),
toolLabels: { coverage_check: 'coverage.check' },
sideEffects: ['order.create'],
});
const deployed = crewaiAdapter({
transport: crewaiHttp({ url: 'https://acme-sales.crewai.com', token: process.env.CREWAI_TOKEN ?? '' }),
});| Option | Meaning |
|---|---|
transport | crewaiProcess(...) or crewaiHttp(...) |
inputs | the crew's inputs for a block, from the history and the turn. Default { message, history, messages }: the block's user text, the earlier turns as role: text lines, and the whole history as JSON. All strings, as CrewAI interpolates them into task descriptions. |
toolLabels, sideEffects, deterministic, version | as on the openai-agents adapter |
crewaiProcess option | Meaning |
|---|---|
crew | the crew, as <module>:<name>: a Crew, or a function that returns a fresh one (one per block, so no state leaks between blocks) |
python, cwd, env, timeoutMs, exec | as on the pydantic-ai adapter |
crewaiHttp option | Meaning |
|---|---|
url, token | the crew's URL and bearer token from the AMP dashboard. Read the token from the environment; it reaches no message. |
pollIntervalMs | the wait between two status reads. Default 2000. |
timeoutMs | the most milliseconds one kickoff may take, polling included. Default 600000. |
fetch, sleep | for tests |
The bridge ​
The adapter writes the bridge (packages/adapters/src/crewai/bridge.py) to the temporary folder and runs python bridge.py <crew>, with { "inputs": {...} } on stdin. The bridge listens on CrewAI's event bus for ToolUsageFinishedEvent and ToolUsageErrorEvent, kicks the crew off, and prints driftbench-result {"raw", "tasks", "tools", "usage"}. A tool that fails and is tried again by CrewAI appears once per attempt.
What send returns ​
Each tool use as a tool call followed by its result (an output that is JSON text is parsed), then the crew's final output as text. Each task's name is a stage, in order. The token usage, when the crew reports it, is the block's cost. The task outputs are kept in raw.tasks.
Capabilities ​
seedPrefix yes (the history goes in the inputs; with the default inputs a Given prefix whose tool calls the spec states as a record, origin: recorded or authored, is refused, since the crew would read them as prose; a canonical prefix, as in mode T, runs); toolStubs no (the crew runs its own tools); endState with sideEffects through the process interface only; the rest no; deterministic as declared.
Errors ​
| What happens | Error |
|---|---|
| a 408, a 429 or a 5xx from AMP (Retry-After kept), or a bridge error with such a status or a connection error | AdapterTransportError: the runner retries under the shared policy |
a crew run that does not finish within timeoutMs | AdapterTransportError, marked exhausted |
any other status, a run whose status is error, or a bridge exit without a result | AgentFailure |
Target ​
The adapter reads no field of the target except retry. Its id is crewai.