Bedrock Agent

Agents, knowledge bases, action groups, flows, prompts, collaborators — full Bedrock Agent control plane.

fakecloud implements the Bedrock Agent control plane: agents, agent aliases and versions, knowledge bases, data sources, ingestion jobs, action groups, agent collaborators, flows (with versions and aliases), prompts, and tag management.

Bedrock Agent is a separate AWS service from Bedrock — it has its own endpoint (bedrock-agent.<region>.amazonaws.com) and its own SDK client (boto3.client("bedrock-agent")). For the model invocation surface, see Bedrock. For the agent invocation surface, see Bedrock Agent Runtime.

Operations

Agents

  • CRUD — CreateAgent, GetAgent, ListAgents, UpdateAgent, DeleteAgent
  • PrepareAgent — transition NOT_PREPARED -> PREPARING -> PREPARED
  • Versions — GetAgentVersion, ListAgentVersions, DeleteAgentVersion
  • Aliases — CreateAgentAlias, GetAgentAlias, ListAgentAliases, UpdateAgentAlias, DeleteAgentAlias

Knowledge bases

  • CRUD — CreateKnowledgeBase, GetKnowledgeBase, ListKnowledgeBases, UpdateKnowledgeBase, DeleteKnowledgeBase
  • Data sources — CreateDataSource, GetDataSource, ListDataSources, UpdateDataSource, DeleteDataSource
  • Ingestion jobs — StartIngestionJob, GetIngestionJob, ListIngestionJobs, StopIngestionJob
  • Documents — DeleteKnowledgeBaseDocuments, GetKnowledgeBaseDocuments, ListKnowledgeBaseDocuments

Action groups

  • CreateAgentActionGroup, GetAgentActionGroup, ListAgentActionGroups, UpdateAgentActionGroup, DeleteAgentActionGroup
  • Supports Lambda executor, OpenAPI schema, function schema, and RETURN_CONTROL action groups
  • Parent action group signatures (AMAZON.UserInput, AMAZON.CodeInterpreter) accepted

Agent <-> knowledge base wiring

  • AssociateAgentKnowledgeBase, DisassociateAgentKnowledgeBase
  • GetAgentKnowledgeBase, ListAgentKnowledgeBases, UpdateAgentKnowledgeBase

Agent collaborators (multi-agent)

  • AssociateAgentCollaborator, DisassociateAgentCollaborator
  • GetAgentCollaborator, ListAgentCollaborators, UpdateAgentCollaborator

Flows

  • CRUD — CreateFlow, GetFlow, ListFlows, UpdateFlow, DeleteFlow, PrepareFlow
  • Versions — CreateFlowVersion, GetFlowVersion, ListFlowVersions, DeleteFlowVersion
  • Aliases — CreateFlowAlias, GetFlowAlias, ListFlowAliases, UpdateFlowAlias, DeleteFlowAlias
  • Validation: ValidateFlowDefinition, with the same findings PrepareFlow records

Prompts

  • CreatePrompt, GetPrompt, ListPrompts, UpdatePrompt, DeletePrompt
  • Variants with model overrides, inference configuration, and template variables

Tags

  • TagResource, UntagResource, ListTagsForResource — on agents, agent aliases, knowledge bases, data sources, flows, flow aliases, prompts

Protocol

REST + JSON. Path-based routing; identifiers are URL-safe strings minted by fakecloud (no AWS-side validation enforced on caller-supplied IDs).

State model

  • Agents start NOT_PREPARED. PrepareAgent advances to PREPARED (and immediately becomes invokable from Bedrock Agent Runtime). DRAFT version is implicit; explicit versions are created on first alias attached to DRAFT.
  • Knowledge bases start CREATING and transition to ACTIVE on next describe. Ingestion jobs follow STARTING -> IN_PROGRESS -> COMPLETE.
  • Data sources transition the same way and persist their vectorIngestionConfiguration verbatim.
  • Flows start NotPrepared. PrepareFlow answers Preparing and validates the definition the way ValidateFlowDefinition does (missing Input/Output nodes, missing or unknown node inputs/outputs, unfed or multiply fed inputs, unknown connection ends, missing connection configuration, incompatible data types, duplicate connections, cycles, unreachable nodes, with AWS's messages and severities). Any Error finding leaves the flow Failed; warnings alone leave it Prepared. GetFlow reports the findings as validations until the draft is updated, which returns it to NotPrepared. Preparing a flow with no definition is a ValidationException. Only a Prepared draft can be versioned (ConflictException otherwise), aliases must route to a numbered version (not DRAFT), a definition holds at most one Input node, and node names must be unique.

Limitations

  • Embedding pipelines are not executed; ingestion jobs report success without indexing real content.
  • Action group Lambda executors and OpenAPI schemas are stored verbatim and returned by the runtime in RETURN_CONTROL traces; the agent does not actually call Lambda for action group execution.
  • Prompt evaluation is not performed; prompts are stored and returned for SDK round-trip tests.
  • Flow validation does not report UnsatisfiedConnectionConditions (conditional paths that can never run), and only checks the required inputs/outputs of Input and Output nodes.

For the data plane (InvokeAgent, Retrieve, RetrieveAndGenerate), see Bedrock Agent Runtime.

Introspection

EndpointMethodDescription
/_fakecloud/bedrock-agent/agentsGETList every agent with aliases, versions, attached knowledge bases, and collaborators flattened into one row each.
curl http://localhost:4566/_fakecloud/bedrock-agent/agents

The response shape is {agents: [{agentId, agentName, agentArn, agentStatus, foundationModel, instruction, knowledgeBases[], actionGroups[], collaborators[], aliases[], versions[], promptOverrides, createdAt, updatedAt}]}. actionGroups is currently always [] because action-group creation does not yet persist state -- the field is exposed for forward compatibility so callers don't have to migrate their assertion code later.

Wrapped by getAgents() on the bedrockAgent sub-client in every first-party SDK.

Source