SYSTEM Cited by 3 sources

Redpanda Agents SDK¶

Redpanda Agents SDK is a three-component preview toolkit released by Redpanda on 2025-04-03 for building enterprise AI agents with an "unreasonably good developer experience" — explicitly marketed as "the Ruby-on-Rails experience for agents". The release coincided with Redpanda's $100M Series D announcement.

Project repo: github.com/redpanda-data/agent.

Three components¶

1. `rpk connect mcp-server`¶

Exposes Redpanda Connect pipelines, resources, and processors as MCP tools. The canonical instantiation of MCP as centralized integration proxy — "allows you to expose any redpanda connect source and destination as a tool with a simple configuration."

Because Redpanda Connect ships ~300 pre-built connectors (databases, queues, caches, SaaS APIs, doc stores, GitHub, Salesforce), enabling MCP for any of them is a config change. See [github.com/redpanda-data/connect/blob/main/internal/plugins/info.csv] for the full connector catalog.

2. Python SDK for agents¶

A Python SDK whose stated focus is:

Durable execution — agent workflows survive process crashes, platform restarts, deploy windows, and resource-limit evictions.
Automatic logs + metrics collection — opt-in full persistence of agent state for debugging, evaluation replay, time-travel debugging.
OpenTelemetry instrumentation — exposed for all MCP pipelines at launch, Python runtime instrumentation shipping after.
Pydantic / OpenAI-agents ergonomic compatibility — "If you have used Pydantic agents or OpenAI agents, the hope is that you will feel right at home."
Redpanda-broker integration — agent-to-agent communication, trace capture, evaluation replay, collaborative threads, message sampling, analytics, explainability-of-actions, time-travel debugging — all backed by Redpanda's distributed log.
BYOC-native — seamless integration with Redpanda's BYOC deployment for sensitive workloads that cannot leave the firewall; identity providers, ACLs, authentication for multi-agent workflows.

3. `rpk connect agent`¶

The glue layer — a CLI that:

Automatically wires MCP tools for your Python agent.
Exposes the Python agent via gRPC local-host command.
Dynamically injects the MCP server address.
Handles tool discovery.
Launches all the tools and Redpanda Connect pipelines.
Exposes the combined surface as simple HTTP endpoints.

The developer workflow Gallego pitches (Source: Gallego 2025-04-03):

# brew upgrade redpanda
# update the connect plugin
rpk connect upgrade

# create a demo repo
mkdir demo && rpk connect agent init .

# profit!
rpk connect agent run

Core pillars (verbatim)¶

From the launch post:

Distributed log — Redpanda storage for durable execution, human-in-the-loop workflows, agent-to-agent communication, trace capture, evaluation replay, logs, metrics, collaborative threads, message sampling, analytics, explainability of actions, time travel debugging, etc.
A new MCP Server project with ~300 pre-built connectors exposes all of your internal tools as a simple HTTP endpoint with a YAML config that manages all of the connection pooling, retries, exponential backoffs, TLS, certificates, authentication, etc.
A Python SDK (in preview) — Pydantic / OpenAI-agents ergonomics, standards-based (OpenTelemetry first), plug-and-play with Redpanda MCP declarative pipelines (rpk connect mcp-server), seamless BYOC integration, identity- provider + ACL + auth for multi-agent workflows.

Starlark as declarative-YAML substitute¶

Because Bloblang is a DSL, Redpanda Agents additionally lets users author Redpanda Connect YAML in Python via the Starlark subset — "effectively Python without imports, but more importantly, it is all Python so no need to learn a new configuration language." This gives the familiar Python experience for teams that prefer code over YAML while keeping the declarative-pipeline benefits for ops.

The "Ruby-on-Rails for agents" framing¶

Gallego's explicit Rails analogy is that the SDK + rpk plumbing is meant to be the conventions-over- configuration layer that turns building an enterprise agent from a bespoke integration project into a rpk connect agent init + code + rpk connect agent run workflow — the same way Rails turned bespoke web-framework assembly into rails new.

Launch scope (2025-04-03)¶

Preview for all Redpanda BYOC customers across three clouds (AWS / GCP / Azure).
Redpanda Serverless support teased for a future update.

Caveats¶

Preview, not GA. The SDK is at preview maturity as of 2025-04-03. Feedback explicitly solicited.
Mechanism depth deferred. Durable-execution guarantees are asserted ("opt-in full persistence") but not described in the source (commit cadence, recovery RPO, replay-correctness model).
OpenTelemetry instrumentation is staged — MCP pipelines instrumented at launch, Python-runtime instrumentation shipping later.
Opinionated Redpanda coupling. The SDK is specifically designed to lean on the Redpanda broker as durable-log substrate — cross-broker portability is not a stated design goal.
Vision-essay context. The launch post frames the SDK as part of a 20-year-arc trajectory (batch → streaming → continuous computation → agentic autonomy); the load-bearing mechanism details for individual components are light on internals.

Productization into ADP (2025-10-28)¶

Seven months after launch, Gallego's 2025-10-28 ADP announcement extends the Agents SDK into an Agentic Data Plane product tier:

Remote MCP + OBO authorization. Extends rpk connect mcp-server local-MCP with remote deployment + on-behalf-of authorization + IdP integration — canonicalised as patterns/on-behalf-of-agent-authorization.
Knowledge-based agent templates — Git, Jira, GDrive out of the box.
Declarative Agent Runtime — opinionated runtime spec above the Python SDK.
Oxla query engine — acquired C++ distributed query engine (PostgreSQL wire, separated compute- storage, Iceberg-native) added as the agent-context-management substrate for SQL-filter-then-model-summarize workloads.
Governance as product surface — access controls + observability elevated from implicit to primary product surface.

The SDK remains the library/SDK layer; ADP is the product-tier composition that packages SDK + streaming + query engine + governance into a single deployment surface across BYOC, Self- Managed, and Cloud.

Seen in¶

sources/2025-10-28-redpanda-introducing-the-agentic-data-plane — productization follow-up; ADP layer.
sources/2025-04-03-redpanda-autonomy-is-the-future-of-infrastructure — the launch post, Alex Gallego's founder-voice announcement of all three components + the Ruby-on-Rails framing + the three-pillar core + the preview-across-three-clouds scope.