Integrations

Concept

Rewyn is meant to sit alongside what you already run, not to replace it. That means two obligations, and they pull in opposite directions.

The core install must stay lean and provider-agnostic, so nothing here is a required dependency and every module imports its SDK lazily. But "we don't depend on your framework" is not the same as "you can use us with your framework", and only the second is useful to someone who already has an agent, a vector database and a search cluster.

So each integration speaks a protocol the SDK already defines. Swapping a backend is a constructor argument, not a migration.

Minimal example

from rewyn.integrations.redis import RedisMemoryStore
from rewyn.memory import Memory

memory = Memory(namespace="tenant:acme", store=RedisMemoryStore("redis://cache:6379/0"))

Production example

Data

from rewyn.integrations.pgvector import PgVectorIndex
from rewyn.integrations.redis import RedisCheckpointStore, RedisMemoryStore
from rewyn.integrations.s3 import S3ObjectStore
from rewyn.integrations.search import SearchRetriever

memory = Memory(namespace=tenant, store=RedisMemoryStore(url))  # shared by every worker
checkpointer = Checkpointer(RedisCheckpointStore(url))  # any worker can resume
pipeline = RAGPipeline(index=PgVectorIndex(connection, dimension=1536))
corpus = SearchRetriever("documents", url="http://opensearch:9200")
objects = S3ObjectStore("rewyn-runs", prefix="prod/")  # for the cloud service

A file-backed store is fine on a laptop and useless behind a load balancer: the second replica cannot read the first one's memory. That is the whole reason Redis is here.

Extras: pip install "rewyn[redis,s3,search]". pgvector needs no extra, because the connection is yours to pool.

Other agent frameworks

Give them your tools, with the schema and validation intact:

from rewyn.integrations.frameworks import to_callables, to_openai_tools

to_openai_tools([search, issue_refund])  # also to_anthropic_tools, to_json_schema_tools
to_callables([search])  # name -> validating callable

Take theirs:

from rewyn.integrations.frameworks import from_schema

weather = from_schema("weather", "Look up the weather.", their_schema, their_function)
agent = Agent(model=..., tools=[weather])

Record theirs:

from rewyn.integrations.frameworks import instrument

graph = instrument(langgraph_app.invoke, framework="langgraph", version="3")
result = graph(question)  # now a recorded Rewyn run

instrument records the input, output, duration and failure. It cannot see inside the other framework, so it does not pretend to. Anything the callable does through Rewyn primitives is recorded in full as usual.

Agent to agent

from rewyn.integrations.a2a import A2AServer, remote_agent_tool

# Serving: framework agnostic, so mount it wherever you already serve.
server = A2AServer(research_agent, endpoint="https://research.internal/a2a")
app.post("/a2a/run")(server.handle)
app.get("/a2a/card")(server.describe)

# Calling: the remote agent is an ordinary tool.
ask_research = remote_agent_tool(
    "https://research.internal/a2a",
    name="ask_research",
    description="Delegate research to the research service.",
)
agent = Agent(model=..., tools=[ask_research])

The call becomes a tool call in the local run, with its own permission check, cost and latency, and the caller's run id crosses the hop so a trace survives the service boundary.

OpenTelemetry

from rewyn.integrations.otel import OTelExporter

with start_run("support") as run:
    OTelExporter().attach(run)

API reference

rewyn/integrations/redis.py, s3.py, search.py, pgvector.py, frameworks.py, a2a.py, otel.py.

Failure modes

MissingDependencyError. The extra is not installed. The message names it.

A backend does not satisfy the protocol. The RAG, memory, sandbox and checkpoint protocols are runtime-checkable, so assert it: assert isinstance(my_index, VectorIndex).

S3 reads return None for everything. A missing object reads as None, but a permissions failure raises. If everything is None, the keys are wrong, not the credentials.

A pgvector insert is rejected. The embedding's length must match the column. The error names both numbers.

An instrumented framework run records nothing inside it. That is expected. Rewyn sees the boundary, not the internals. Use Rewyn primitives inside the callable if you want the detail.

A remote agent call fails silently. It does not: A2AResponse.ok is false and error says why. Through remote_agent_tool it raises, which the agent loop records as a tool error.