Environment Variables

The AgentCore adapter injects environment variables into agent runtimes via the EnvironmentVariables field on CreateAgentRuntime and UpdateAgentRuntime API calls. These variables allow runtime code to discover configuration, peer agents, and deployed resources.

Variable reference

Variable	Source	When Set	Description
PROMPTPACK_PROVIDER_TYPE	Arena config deploy.agentcore	Always (code deploy)	LLM provider type (e.g. "bedrock"). Used by the runtime to select the correct provider.
PROMPTPACK_PROVIDER_MODEL	Arena config deploy.agentcore.model	Always (code deploy)	Bedrock model ID (e.g. "claude-3-5-haiku-20241022"). Used by the runtime to configure the LLM.
PROMPTPACK_PACK_JSON	Pack file contents	Always (code deploy)	The full pack JSON, injected so the runtime can load the pack without a separate file.
PROMPTPACK_LOG_GROUP	observability.cloudwatch_log_group	When cloudwatch_log_group is a non-empty string	CloudWatch log group name for structured logging.
PROMPTPACK_TRACING_ENABLED	observability.tracing_enabled	When tracing_enabled is true	Enables AWS X-Ray tracing. Value is the string "true".
PROMPTPACK_MEMORY_STORE	memory_store config field	When memory_store is set	Memory store type: "session" or "persistent".
PROMPTPACK_MEMORY_ID	Memory resource ARN	After memory resource creation during Apply	The ARN of the created memory resource. Allows runtimes to connect to the memory store.
PROMPTPACK_AGENTS	Runtime resource ARNs	Multi-agent packs only, after all runtimes are created	JSON object mapping agent member names to their runtime ARNs. Injected on the entry agent only.
PROMPTPACK_A2A_AUTH_MODE	a2a_auth.mode	When a2a_auth is configured with a non-empty mode	A2A authentication mode: "iam" or "jwt".
PROMPTPACK_A2A_AUTH_ROLE	runtime_role_arn	When a2a_auth.mode is "iam"	The IAM role ARN used for A2A authentication between agents.
PROMPTPACK_POLICY_ENGINE_ARN	Cedar policy resource ARNs	After Cedar policy creation during Apply	Comma-separated list of policy engine ARNs. Set when prompts define validators or tool_policy.
PROMPTPACK_METRICS_CONFIG	Pack evals with metrics	When at least one eval defines a metric	JSON MetricsConfig object describing CloudWatch metrics for eval reporting.
PROMPTPACK_DASHBOARD_CONFIG	Pack structure (agents + evals)	When the pack has agents or eval metrics	JSON DashboardConfig object describing a CloudWatch dashboard layout.
PROMPTPACK_PROTOCOL	protocol config field	When protocol is set to a non-empty value	Server protocol mode: "http", "a2a", or "both". Controls which servers the runtime starts. See Runtime Protocols.
PROMPTPACK_AGENT	Pack prompt/agent name	Multi-agent packs; set per-runtime	The agent name this runtime serves. Omitted for single-agent packs to allow auto-discovery.

Variable details

PROMPTPACK_PROVIDER_TYPE

Set from the arena config’s deploy.agentcore section. Tells the runtime which LLM provider to use. For AgentCore deployments, this is always "bedrock".

PROMPTPACK_PROVIDER_TYPE=bedrock

PROMPTPACK_PROVIDER_MODEL

Set from the arena config’s deploy.agentcore.model field. Specifies the Bedrock model ID the runtime should use for LLM invocations.

PROMPTPACK_PROVIDER_MODEL=claude-3-5-haiku-20241022

PROMPTPACK_PACK_JSON

Injected during code deploy. Contains the entire compiled pack JSON so the runtime can load the pack directly from the environment without needing a separate file on disk.

PROMPTPACK_PACK_JSON={"id":"my-agent","version":"v1.0.0","prompts":{...}}

PROMPTPACK_LOG_GROUP

Set from observability.cloudwatch_log_group. Runtimes use this to direct structured logs to the specified CloudWatch log group.

PROMPTPACK_LOG_GROUP=/aws/agentcore/my-pack

PROMPTPACK_TRACING_ENABLED

Set from observability.tracing_enabled. Only injected when the value is true. Runtimes use this to enable X-Ray trace instrumentation.

PROMPTPACK_TRACING_ENABLED=true

PROMPTPACK_MEMORY_STORE

Set from the top-level memory_store config field. Tells the runtime which memory strategy to use.

PROMPTPACK_MEMORY_STORE=session

PROMPTPACK_MEMORY_ID

Injected at Apply time after the memory resource is successfully created. Contains the full ARN of the memory resource.

PROMPTPACK_MEMORY_ID=arn:aws:bedrock:us-west-2:123456789012:memory/abc123

PROMPTPACK_AGENTS

Injected on the entry agent only in multi-agent packs, after all agent runtimes are created. Contains a JSON object mapping each agent member name to its runtime ARN.

{
  "planner": "arn:aws:bedrock:us-west-2:123456789012:agent-runtime/rt-001",
  "researcher": "arn:aws:bedrock:us-west-2:123456789012:agent-runtime/rt-002",
  "writer": "arn:aws:bedrock:us-west-2:123456789012:agent-runtime/rt-003"
}

Only runtimes with status "created" or "updated" are included in the map. If a member runtime failed to create, it is omitted.

PROMPTPACK_A2A_AUTH_MODE

Set when a2a_auth is configured. Indicates the authentication mechanism used for agent-to-agent communication.

PROMPTPACK_A2A_AUTH_MODE=iam

PROMPTPACK_A2A_AUTH_ROLE

Set when a2a_auth.mode is "iam" and runtime_role_arn is non-empty. Provides the IAM role ARN that agents use to authenticate with each other.

PROMPTPACK_A2A_AUTH_ROLE=arn:aws:iam::123456789012:role/AgentCoreRuntime

PROMPTPACK_POLICY_ENGINE_ARN

Injected after all Cedar policy resources are created. Contains a comma-separated list of policy engine ARNs, one per prompt that has validators or tool_policy.

PROMPTPACK_POLICY_ENGINE_ARN=arn:aws:bedrock:us-west-2:123456789012:policy-engine/pe-001,arn:aws:bedrock:us-west-2:123456789012:policy-engine/pe-002

PROMPTPACK_METRICS_CONFIG

Injected when at least one eval in the pack defines a metric. Contains a JSON MetricsConfig object that describes the CloudWatch metrics the runtime should emit.

Example payload:

{
  "namespace": "PromptPack/Evals",
  "dimensions": {
    "pack_id": "my-pack",
    "agent": "multi"
  },
  "metrics": [
    {
      "eval_id": "accuracy_eval",
      "metric_name": "accuracy_score",
      "metric_type": "gauge",
      "unit": "None"
    },
    {
      "eval_id": "latency_eval",
      "metric_name": "response_latency",
      "metric_type": "histogram",
      "unit": "Milliseconds"
    },
    {
      "eval_id": "error_eval",
      "metric_name": "error_count",
      "metric_type": "counter",
      "unit": "Count"
    }
  ],
  "alarms": [
    {
      "metric_name": "accuracy_score",
      "min": 0.8
    },
    {
      "metric_name": "response_latency",
      "max": 5000.0
    }
  ]
}

MetricsConfig fields:

Field	Type	Description
namespace	string	CloudWatch namespace. Always "PromptPack/Evals".
dimensions	map[string]string	Dimension key-value pairs. Always includes pack_id. Includes "agent": "multi" for multi-agent packs.
metrics	MetricEntry[]	One entry per eval metric.
alarms	AlarmEntry[]	Optional. One entry per metric that defines a range.

MetricEntry fields:

Field	Type	Description
eval_id	string	The eval identifier from the pack definition.
metric_name	string	CloudWatch metric name.
metric_type	string	Metric type from the eval definition (e.g. "gauge", "counter", "histogram", "boolean").
unit	string	CloudWatch unit. Mapped from metric type: counter -> "Count", histogram -> "Milliseconds", gauge/boolean -> "None".

AlarmEntry fields:

Field	Type	Description
metric_name	string	The metric this alarm applies to.
min	float64 (optional)	Minimum acceptable value.
max	float64 (optional)	Maximum acceptable value.

PROMPTPACK_DASHBOARD_CONFIG

Injected when the pack has agents or eval metrics. Contains a JSON DashboardConfig object that describes a CloudWatch dashboard layout with widgets for agent runtime metrics, A2A latency (multi-agent), and eval metrics.

Example payload (multi-agent pack with one eval metric):

{
  "widgets": [
    {
      "type": "metric",
      "x": 0,
      "y": 0,
      "width": 12,
      "height": 6,
      "properties": {
        "title": "Agent: planner",
        "region": "us-west-2",
        "period": 300,
        "metrics": [
          ["PromptPack/Evals", "Invocations", "agent", "planner"],
          ["PromptPack/Evals", "Errors", "agent", "planner"],
          ["PromptPack/Evals", "Duration", "agent", "planner"]
        ]
      }
    },
    {
      "type": "metric",
      "x": 12,
      "y": 0,
      "width": 12,
      "height": 6,
      "properties": {
        "title": "Agent: researcher",
        "region": "us-west-2",
        "period": 300,
        "metrics": [
          ["PromptPack/Evals", "Invocations", "agent", "researcher"],
          ["PromptPack/Evals", "Errors", "agent", "researcher"],
          ["PromptPack/Evals", "Duration", "agent", "researcher"]
        ]
      }
    },
    {
      "type": "metric",
      "x": 0,
      "y": 6,
      "width": 24,
      "height": 6,
      "properties": {
        "title": "Inter-Agent A2A Call Latency",
        "region": "us-west-2",
        "period": 300,
        "metrics": [
          ["PromptPack/Evals", "A2ALatency", "agent", "planner"],
          ["PromptPack/Evals", "A2ALatency", "agent", "researcher"]
        ]
      }
    },
    {
      "type": "metric",
      "x": 0,
      "y": 12,
      "width": 12,
      "height": 6,
      "properties": {
        "title": "Eval: accuracy_score",
        "region": "us-west-2",
        "period": 300,
        "metrics": [
          ["PromptPack/Evals", "accuracy_score", "pack_id", "my-pack"]
        ],
        "annotations": {
          "horizontal": [
            { "label": "min", "value": 0.8, "color": "#2ca02c" },
            { "label": "max", "value": 1.0, "color": "#d62728" }
          ]
        }
      }
    }
  ]
}

DashboardConfig fields:

Field	Type	Description
widgets	DashboardWidget[]	Ordered list of dashboard widgets.

DashboardWidget fields:

Field	Type	Description
type	string	Always "metric".
x	int	Horizontal grid position (0 or 12 for two-column layout).
y	int	Vertical grid position. Increments by 6 (widget height) per row.
width	int	Widget width. 12 for standard widgets, 24 for full-width (A2A latency).
height	int	Widget height. Always 6.
properties	object	Widget display properties including title, region, period, metrics, and optional annotations.

Widget types generated:

Widget	When Generated	Layout
Agent runtime metrics (Invocations, Errors, Duration)	One per agent member (or pack ID for single-agent)	Two-column, 12 units wide
A2A call latency	Multi-agent packs only	Full-width, 24 units wide
Eval metric	One per eval with a metric definition	Two-column, 12 units wide, with optional min/max threshold annotations

PROMPTPACK_PROTOCOL

Set from the protocol config field. Controls which servers the runtime starts: the HTTP bridge (port 8080), the A2A server (port 9000), or both.

PROMPTPACK_PROTOCOL=both

See Runtime Protocols for details on the HTTP bridge endpoints and payload formats.

PROMPTPACK_AGENT

Set per-runtime in multi-agent packs. Contains the agent/prompt name this runtime serves. Omitted for single-agent packs to allow the runtime to auto-discover the single prompt from the pack.

PROMPTPACK_AGENT=coordinator

Injection timing

Environment variables are built and injected at different points during the Apply lifecycle:

Timing	Variables
Before any resource creation	PROMPTPACK_PROVIDER_TYPE, PROMPTPACK_PROVIDER_MODEL, PROMPTPACK_PACK_JSON, PROMPTPACK_LOG_GROUP, PROMPTPACK_TRACING_ENABLED, PROMPTPACK_MEMORY_STORE, PROMPTPACK_A2A_AUTH_MODE, PROMPTPACK_A2A_AUTH_ROLE, PROMPTPACK_METRICS_CONFIG, PROMPTPACK_DASHBOARD_CONFIG, PROMPTPACK_PROTOCOL, PROMPTPACK_AGENT
After memory creation (pre-step)	PROMPTPACK_MEMORY_ID
After Cedar policy creation (phase 2)	PROMPTPACK_POLICY_ENGINE_ARN
After runtime creation (phase 3)	PROMPTPACK_AGENTS (injected via UpdateRuntime on entry agent)

Variables injected before resource creation are available to all runtimes at creation time. Variables injected after a phase require a subsequent UpdateAgentRuntime call to propagate to already-created runtimes.