Agentic Data Plane: Governance
The Agentic Data Plane (ADP) decides whether an agent's call runs (access policies, roles, the
kill switch), what it may cost (budgets), what data it exposes (guardrails, data policies), and
which credentials it uses (secrets, API keys, OAuth). Sign in and pick an environment first, as
/redpanda:adp describes. This skill keeps only the procedures below. Look up everything else in
the docs (last section).
How to work
Before you delete a secret, find what still uses it. rpk ai secret delete checks no
references, and a secret's value can't be read back, so a deleted secret can't be restored and
every resource that names it fails at its next call. Read the MCP servers, LLM providers, OAuth
providers and clients, and agents (list -o yaml) for the secret's ID first. If anything still
uses it, tell the user what, and ask before deleting.
To deny one tool on one MCP server, name the tool alone and pin the server in a when clause.
Cedar identifies a tool by its name only, so write resource == McpServerTool::"<tool>" with
when { resource in McpServer::"<server>" }. A combined ID such as
McpServerTool::"<server>/<tool>" names no tool, so it denies nothing, and leaving out the
when clause denies that tool name on every server.
To stop a misbehaving agent, contain it. Don't delete it, its credentials, or its MCP servers.
Containment (rpk ai agent-kill-switch, or Contain agent in the UI, with a reason) denies the
agent's gateway traffic, to it and from it, until someone releases it, and leaves its
configuration as it is.
Deletes can't be undone and remove what an investigation needs.
Use the docs for everything else
Look up commands, flags, Cedar actions and attributes, permissions, templates, limits, and UI
steps in the Redpanda docs instead of from memory:
- Follow
/redpanda:redpanda-docs. The ADP docs are not versioned and not tied to a platform, so
ask_redpanda_question needs no version. Use get_rpk_command for one rpk ai command and
list_rpk_commands with plugin: "ai" to find one.
- Most useful pages, under https://docs.redpanda.com/agentic-data-plane/:
control/access-policies/
(Cedar, templates, the action reference), control/permissions-overview/ and
control/permissions-reference/ (roles and permissions), control/budgets/,
control/cost-usage/, control/guardrails/, control/kill-switch/, control/secrets/,
control/tags/, connect/data-policies/, connect/oauth-providers/,
connect/remote-mcp-clients/, and the rpk ai reference under reference/rpk/rpk-ai/.
If the Docs MCP server refuses for quota and you can't fetch the page, tell the user to sign in
through /mcp rather than answering from memory.
rpk ai <command> --help on the installed plugin is the live check for flags in that version.
1---2name: adp-governance3description: >-4 Govern what Agentic Data Plane agents may do, spend, and expose, through the `rpk ai` CLI and the5 ADP UI at ai.redpanda.com, and look up governance features in the Redpanda docs. Use when writing6 or debugging a Cedar access policy with `rpk ai policy`, giving a new agent access, capping agent7 spend with `rpk ai budget`, analyzing cost, attaching a guardrail to a provider, containing or8 releasing an agent with the kill switch, storing or deleting a credential with `rpk ai secret`,9 minting an API key, masking MCP tool data with data policies, or setting up inbound or outbound10 OAuth. For signing in, agents, MCP servers, LLM providers, and the rest of the `rpk ai` CLI, see11 `/redpanda:adp`. For transcripts, the audit log, and traces, see `/redpanda:adp-observability`.12metadata:13 source: https://skillmd.com/skills/redpanda-data/adp-governance14---1516# Agentic Data Plane: Governance1718The Agentic Data Plane (ADP) decides whether an agent's call runs (access policies, roles, the19kill switch), what it may cost (budgets), what data it exposes (guardrails, data policies), and20which credentials it uses (secrets, API keys, OAuth). Sign in and pick an environment first, as21`/redpanda:adp` describes. This skill keeps only the procedures below. Look up everything else in22the docs (last section).2324## How to work2526**Before you delete a secret, find what still uses it.** `rpk ai secret delete` checks no27references, and a secret's value can't be read back, so a deleted secret can't be restored and28every resource that names it fails at its next call. Read the MCP servers, LLM providers, OAuth29providers and clients, and agents (`list -o yaml`) for the secret's ID first. If anything still30uses it, tell the user what, and ask before deleting.3132**To deny one tool on one MCP server, name the tool alone and pin the server in a `when` clause.**33Cedar identifies a tool by its name only, so write `resource == McpServerTool::"<tool>"` with34`when { resource in McpServer::"<server>" }`. A combined ID such as35`McpServerTool::"<server>/<tool>"` names no tool, so it denies nothing, and leaving out the36`when` clause denies that tool name on every server.3738**To stop a misbehaving agent, contain it. Don't delete it, its credentials, or its MCP servers.**39Containment (`rpk ai agent-kill-switch`, or **Contain agent** in the UI, with a reason) denies the40agent's gateway traffic, to it and from it, until someone releases it, and leaves its41configuration as it is.42Deletes can't be undone and remove what an investigation needs.4344## Use the docs for everything else4546Look up commands, flags, Cedar actions and attributes, permissions, templates, limits, and UI47steps in the Redpanda docs instead of from memory:4849- Follow `/redpanda:redpanda-docs`. The ADP docs are not versioned and not tied to a platform, so50 `ask_redpanda_question` needs no `version`. Use `get_rpk_command` for one `rpk ai` command and51 `list_rpk_commands` with `plugin: "ai"` to find one.52- Most useful pages, under https://docs.redpanda.com/agentic-data-plane/: `control/access-policies/`53 (Cedar, templates, the action reference), `control/permissions-overview/` and54 `control/permissions-reference/` (roles and permissions), `control/budgets/`,55 `control/cost-usage/`, `control/guardrails/`, `control/kill-switch/`, `control/secrets/`,56 `control/tags/`, `connect/data-policies/`, `connect/oauth-providers/`,57 `connect/remote-mcp-clients/`, and the `rpk ai` reference under `reference/rpk/rpk-ai/`.5859If the Docs MCP server refuses for quota and you can't fetch the page, tell the user to sign in60through `/mcp` rather than answering from memory.6162`rpk ai <command> --help` on the installed plugin is the live check for flags in that version.