Resource Types

The AgentCore adapter manages seven resource types. Each resource has a constant name used in state serialization, a mapping to the PromptPack concept it represents, and defined create/update/delete/health-check behavior.

Resource type summary

Constant	String Value	Pack Concept	Create	Update	Delete	Health Check
ResTypeMemory	memory	Memory store config	Yes	No	Yes	Status ACTIVE
ResTypeToolGateway	tool_gateway	Pack tools	Yes	No	Yes	Status READY
ResTypeCedarPolicy	cedar_policy	Prompt validators / tool_policy	Yes	No	Yes	Engine ACTIVE
ResTypeAgentRuntime	agent_runtime	Agent members (or pack ID)	Yes	Yes	Yes	Status READY
ResTypeA2AEndpoint	a2a_endpoint	Multi-agent wiring	Yes	No	No-op	Always healthy
ResTypeEvaluator	evaluator	Pack evals (llm_as_judge only)	Yes	No	Yes	Status ACTIVE
ResTypeOnlineEvalConfig	online_eval_config	Wires evaluators to agent traces	Yes	No	Yes	Status ACTIVE

Resource status values

Resources pass through these status values during their lifecycle:

Status	Constant	Meaning
created	ResStatusCreated	Resource was successfully created during Apply.
updated	ResStatusUpdated	Resource was successfully updated during Apply (redeployment).
failed	ResStatusFailed	Resource creation or update failed. The error is reported via callback.
planned	ResStatusPlanned	Resource would be created (dry-run mode only).

Health check status values returned by Status:

Status	Constant	Meaning
healthy	StatusHealthy	Resource exists and is in its expected ready/active state.
unhealthy	StatusUnhealthy	Resource exists but is not in the expected state, or the check returned an error.
missing	StatusMissing	Resource was not found (404/NotFound from AWS).

---

memory

Constant: ResTypeMemory String value: "memory"

Pack mapping

Created when memory_store is set in the deploy config. The resource name is {pack_id}_memory. One memory resource is created per pack.

AWS API calls

Operation	API Call	Details
Create	CreateMemory	Provisions a Bedrock AgentCore memory with the configured strategy (episodic for "session", semantic for "persistent"). Sets event expiry to 30 days.
Delete	DeleteMemory	Deletes the memory resource by ID. Tolerates NotFound (already deleted).

Health check

Calls GetMemory and checks that Memory.Status equals ACTIVE.

Result	Condition
healthy	Status is ACTIVE
unhealthy	Status is any other value, or API error
missing	NotFound error

Side effects

On successful creation, the memory ARN is injected into PROMPTPACK_MEMORY_ID on the runtime config so that agent runtimes can discover the memory resource.

---

tool_gateway

Constant: ResTypeToolGateway String value: "tool_gateway"

Pack mapping

One tool_gateway resource is created per entry in pack.Tools. Resources are created in sorted key order.

AWS API calls

Operation	API Call	Details
Create (parent)	CreateGateway	Lazily creates a shared parent gateway on the first tool. The gateway uses MCP protocol type and no authorizer. Polls until READY.
Create (target)	CreateGatewayTarget	Creates a gateway target for each tool within the shared gateway.
Delete	DeleteGateway	Deletes the parent gateway by ID. Tolerates NotFound.

The parent gateway is created lazily on the first CreateGatewayTool call and reused for all subsequent targets within the same Apply invocation. The gateway name is {first_tool_name}_gw.

Health check

Calls GetGateway and checks that Status equals READY.

Result	Condition
healthy	Status is READY
unhealthy	Status is any other value, or API error
missing	NotFound error

Update support

Not supported. Redeployment creates new gateway targets.

---

cedar_policy

Constant: ResTypeCedarPolicy String value: "cedar_policy"

Pack mapping

One cedar_policy resource is created per prompt that has validators or tool_policy defined. The adapter generates Cedar policy statements from these definitions.

AWS API calls

Operation	API Call	Details
Create (engine)	CreatePolicyEngine	Creates a policy engine per prompt. Polls until engine status is ACTIVE.
Create (policy)	CreatePolicy	Creates a Cedar policy within the engine using the generated statement.
Delete (policy)	DeletePolicy	Deletes the Cedar policy by engine ID and policy ID. Tolerates NotFound.
Delete (engine)	DeletePolicyEngine	Deletes the policy engine by ID. Tolerates NotFound.

Health check

Calls GetPolicyEngine and checks that Status equals ACTIVE.

Result	Condition
healthy	Engine status is ACTIVE
unhealthy	Engine status is any other value, or API error
missing	NotFound error, or no policy_engine_id in metadata

Metadata

The resource state stores additional metadata used for deletion and health checks:

Key	Description
policy_engine_id	The policy engine identifier.
policy_engine_arn	The policy engine ARN. Used to populate PROMPTPACK_POLICY_ENGINE_ARN.
policy_id	The Cedar policy identifier within the engine.

Side effects

After all policy resources are created, the adapter injects PROMPTPACK_POLICY_ENGINE_ARN into the runtime config as a comma-separated list of engine ARNs.

Update support

Not supported.

---

agent_runtime

Constant: ResTypeAgentRuntime String value: "agent_runtime"

Pack mapping

For multi-agent packs, one runtime is created per agent member (using the agent name). For single-agent packs, one runtime is created using the pack ID.

AWS API calls

Operation	API Call	Details
Create	CreateAgentRuntime	Provisions an AgentCore runtime with the configured role, environment variables, authorizer, and tags. Polls until status is READY.
Update	UpdateAgentRuntime	Updates an existing runtime with new environment variables and authorizer config. Polls until status is READY. Triggered on redeployment when the resource exists in prior state.
Delete	DeleteAgentRuntime	Deletes the runtime by ID. Tolerates NotFound.

This is the only resource type that supports update. On redeployment, if a runtime with the same type and name exists in the prior state, the adapter calls UpdateAgentRuntime instead of CreateAgentRuntime.

Health check

Calls GetAgentRuntime and checks that Status equals READY.

Result	Condition
healthy	Status is READY
unhealthy	Status is CREATE_FAILED, UPDATE_FAILED, or any non-READY value
missing	NotFound error

Polling behavior

After creation or update, the adapter polls GetAgentRuntime every 5 seconds for up to 60 attempts (5 minutes). Terminal failure states (CREATE_FAILED, UPDATE_FAILED) abort polling immediately. The runtime ARN is returned even if polling fails, allowing the state to record a partial result.

Side effects

For multi-agent packs, after all runtimes are created, the adapter builds a JSON map of {agentName: runtimeARN} and injects it as PROMPTPACK_AGENTS on the entry agent via an UpdateAgentRuntime call.

---

a2a_endpoint

Constant: ResTypeA2AEndpoint String value: "a2a_endpoint"

Pack mapping

One a2a_endpoint resource is created per agent member in multi-agent packs. The resource name is {agent_name}_a2a. Not created for single-agent packs.

AWS API calls

This is a logical resource. No separate AWS API call is made. The AgentCore runtime exposes A2A endpoints when configured with the appropriate environment variables.

Operation	API Call	Details
Create	None	Returns a placeholder ARN: arn:aws:bedrock:{region}:a2a-endpoint/{name}
Delete	No-op	Logged and skipped.

Health check

Always returns healthy. No AWS API call is made.

Update support

Not supported.

---

evaluator

Constant: ResTypeEvaluator String value: "evaluator"

Pack mapping

One evaluator resource is created per llm_as_judge eval in pack.Evals. Other eval types (regex, contains, etc.) are local-only and do not create AWS resources. The resource name is the eval’s ID field, or eval_{index} if the ID is empty.

AWS API calls

Operation	API Call	Details
Create	CreateEvaluator	Provisions an LLM-as-a-Judge evaluator with instructions, model config, and a numerical rating scale. Polls until status is ACTIVE.
Delete	DeleteEvaluator	Deletes the evaluator by ID. Tolerates NotFound (already deleted).

The eval definition’s trigger field maps to the SDK evaluator level: every_turn and sample_turns map to TRACE, while on_session_complete and sample_sessions map to SESSION.

The params map supports the following keys:

Key	Default	Description
instructions	"Evaluate the agent response quality."	Evaluation instructions for the LLM judge.
model	anthropic.claude-sonnet-4-20250514-v1:0	Bedrock model ID for evaluation.
rating_scale_size	5	Number of levels in the numerical 1–N rating scale.

Health check

Calls GetEvaluator and checks that Status equals ACTIVE.

Result	Condition
healthy	Status is ACTIVE
unhealthy	Status is any other value, or API error
missing	NotFound error

Polling behavior

After creation, the adapter polls GetEvaluator every 5 seconds for up to 60 attempts (5 minutes). Terminal failure states (CREATE_FAILED, UPDATE_FAILED) abort polling immediately.

Update support

Not supported.

---

online_eval_config

Constant: ResTypeOnlineEvalConfig String value: "online_eval_config"

Pack mapping

One online_eval_config resource is created per pack when the pack has any llm_as_judge evals. The resource name is {pack_id}_online_eval. It wires the evaluators created in the previous phase to agent runtime traces via CloudWatch logs.

AWS API calls

Operation	API Call	Details
Create	CreateOnlineEvaluationConfig	Creates an online evaluation config referencing all evaluator IDs, a CloudWatch data source, and a sampling rule. Polls until status is ACTIVE.
Delete	DeleteOnlineEvaluationConfig	Deletes the config by ID. Tolerates NotFound (already deleted).

The CloudWatch log group is resolved from observability.cloudwatch_log_group if configured, otherwise defaults to /aws/bedrock/agentcore/{pack_id}. The sampling percentage defaults to 100% but can be overridden via the sample_percentage eval param.

Health check

Calls GetOnlineEvaluationConfig and checks that Status equals ACTIVE.

Result	Condition
healthy	Status is ACTIVE
unhealthy	Status is any other value, or API error
missing	NotFound error

Polling behavior

After creation, the adapter polls GetOnlineEvaluationConfig every 5 seconds for up to 60 attempts (5 minutes). Terminal failure states (CREATE_FAILED, UPDATE_FAILED) abort polling immediately.

Update support

Not supported.

---

Deploy phase ordering

Resources are created during Apply in dependency order across six phases:

Phase	Step Index	Resource Type	Progress Range
Pre-step	—	memory	0%
1	0	tool_gateway	0—17%
2	1	cedar_policy	17—33%
3	2	agent_runtime	33—50%
4	3	a2a_endpoint	50—67%
5	4	evaluator	67—83%
6	5	online_eval_config	83—100%

Destroy ordering

Resources are destroyed in reverse dependency order:

1. online_eval_config
2. cedar_policy
3. evaluator
4. a2a_endpoint
5. agent_runtime
6. tool_gateway
7. memory

Any resource types not in this list are destroyed last, after the ordered groups.