feat(aws): agentcore harness adapter by tshepomaredi · Pull Request #7287 · camunda/connectors

tshepomaredi · 2026-05-20T07:00:28Z

Description

This PR adds AWS Bedrock AgentCore Harness integration to the agentic-ai framework, enabling BPMN ad-hoc subprocesses to use AWS-managed AI agents (AgentCore Harness) with full visibility of custom in-line tool calls as BPMN element activations.

Key features

Introduces AgentCoreHarnessAdapter implementing AiFrameworkAdapter for Harness API integration
Adds "AgentCore Managed Agent" element template for Camunda Modeler
Inner BPMN elements become inline_function tools, giving Camunda control over tool execution
Built-in Harness tools (shell, browser, code interpreter) can be used alongside custom BPMN tools
Supports allowedTools configuration to scope which tools the agent can use

Related issues

closes #

Checklist

Backport labels are added if these code changes should be backported. No backport label is added to the latest
release, as this branch will be rebased onto main before the next release. Example backport labels:
- backport stable/8.8: for changes that should be included in the next 8.8.x release.
- or backport release-8.8.7: for changes that should be included in the specific release 8.8.7, and this
  release has already been created. The release branch will be merged back into stable/8.8 later, so the change
  will be included in future 8.8.x releases as well.
Tests/Integration tests for the changes have been added if applicable.
If the change requires a documentation update, it has been added to the appropriate section in the documentation.

Add initial implementation of AgentCoreHarnessAdapter that implements AiFrameworkAdapter interface for AWS Bedrock AgentCore Harness integration. Components: - AgentCoreHarnessAiFrameworkChatResponse: Response record with session ID - HarnessMessageConverter: Converts Camunda messages to/from Harness format - HarnessToolConverter: Converts ToolDefinition to HarnessTool (inline_function) - AgentCoreHarnessAdapter: Main adapter calling InvokeHarness API The adapter uses inline_function tools so Harness returns tool calls back to Camunda for BPMN element activation, preserving visibility and control.

- Add AgentCoreHarnessJobWorker and AgentCoreHarnessRequest for ad-hoc subprocess - Add element template for 'AgentCore Managed Agent' in modeler - Wire up Spring bean configuration in AgenticAiConnectorsAutoConfiguration - Handle stopReason from Harness response (tool_use vs end_turn) - Filter tool calls to only activate BPMN elements for inline_function tools - Built-in tools (shell, browser) are executed by Harness internally - Add allowedTools support to scope which tools Harness can use - Add INFO logging for debugging Harness invocations

CLAassistant · 2026-05-20T07:05:27Z

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

+  }
+
+  /** Accumulates data from streaming Harness response events. */
+  private class HarnessResponseData {


+      textBuilder.append(text);
+    }
+
+    void startToolUse(HarnessContentBlock.Builder toolUseBuilder) {


- Add HarnessToolConverterTest (7 tests) - Add HarnessMessageConverterTest (12 tests) - Cover tool definition conversion, message conversion, and error handling

nikonovd · 2026-05-27T11:55:24Z

@@ -0,0 +1,247 @@
+{
+  "$schema": "https://unpkg.com/@camunda/zeebe-element-templates-json-schema/resources/schema.json",
+  "name": "AgentCore Managed Agent",


❓ should we name it Amazon Bedrock AgentCore Managed Agent?

we should also highlight it as an Alpha Feature while the AgentCore Harness itself is still in Preview.

nikonovd · 2026-05-27T11:55:53Z

+  "name": "AgentCore Managed Agent",
+  "id": "io.camunda.connectors.agenticai.agentcore.harness.v1",
+  "description": "Drive an AWS Bedrock AgentCore Harness through a re-entrant tool-call loop, exposing the AHSP inner elements as inline_function tools.",
+  "keywords": ["AI", "AI Agent", "AWS", "AgentCore", "Harness", "Bedrock"],


Suggested change

"keywords": ["AI", "AI Agent", "AWS", "AgentCore", "Harness", "Bedrock"],

"keywords": ["AI", "AI Agent", "Amazon", "AWS", "AgentCore", "Harness", "Bedrock"],

🔧 we should add Amazon as a keyword

nikonovd · 2026-05-27T11:56:18Z

+    "value": "bpmn:AdHocSubProcess"
+  },
+  "engines": {
+    "camunda": "^8.9"


Suggested change

"camunda": "^8.9"

"camunda": "^8.10"

nikonovd · 2026-05-27T11:56:47Z

                LOGGER.trace("Activating tool {}: {}", toolCall.metadata().name(), toolCall);
              } else {
-                LOGGER.debug("Activating tool {}", toolCall.metadata().name());
+                LOGGER.info("Activating tool {}", toolCall.metadata().name());


❓ do we need that log on the info level?

Reverted to debug. The info level was added during development for debugging

nikonovd · 2026-05-27T11:57:18Z

+      AgentCoreHarnessJobWorker.USER_PROMPT_VARIABLE,
+      AgentCoreHarnessJobWorker.MAX_ITERATIONS_VARIABLE
+    })
+public class AgentCoreHarnessJobWorker implements AgentConnectorFunction {


🔧 we should actually start to call it AgentCoreHarnessSubProcess to begin getting rid of the JobWorker suffix.

✅ . Renamed to AgentCoreHarnessSubProcess and updated constants from JOB_WORKER_NAME/TYPE to CONNECTOR_NAME/TYPE

nikonovd · 2026-05-27T12:21:46Z

+        agentContext,
+        toolCallResults,
+        null, // provider is not used for Harness
+        new JobWorkerAgentRequestData(


💭 I think this shows that we need to adjust the JobWorkerExecutionContext to maybe have a different type of AgentRequest that it can carry, as currently you are glueing here a lot of data to a completely different execution model.

True, this is a workaround to reuse the existing JobWorkerAgentRequestHandler. A cleaner approach would be to introduce a more generic AgentExecutionContext, or something that doesn't assume the Langchain provider model. I'd suggest we track this as a follow-up refactor to keep this PR focused on the Harness integration. WDYT?

nikonovd · 2026-05-27T12:25:45Z

+      @NotBlank String harnessArn, String region, List<String> allowedTools) {
+
+    /** Extract region from ARN if not explicitly provided. */
+    public String effectiveRegion() {


💭 Nice feature, but currently i think the region is an explicity and mandatory config param, which we should also enforce in such a case.

The region inference from ARN was added for convenience since the ARN always contains the region (arn:aws:bedrock-agentcore:REGION:ACCOUNT:harness/NAME). But I can make it mandatory for consistency with other AWS connectors.

nikonovd · 2026-05-27T12:27:57Z

+        updatedAgentContext, assistantMessage, sessionId);
+  }
+
+  private String getOrCreateSessionId(AgentContext agentContext) {


❓ Is the session subject to timeouts? Can we restore sessions or is it eternal until we close it?

Per AWS documentation, the runtimeSessionId is used for conversation continuity within a Harness invocation session. The session is managed by Harness and persists across multiple InvokeHarness calls with the same session ID. I'll add a comment documenting this behavior. Regarding timeouts, this depends on the Harness config AWS, the session ID itself doesn't expire but the underlying Harness session may have idle timeouts configured.

nikonovd · 2026-05-27T12:31:40Z

+    return responseData;
+  }
+
+  private AssistantMessage buildAssistantMessage(


🔧 Likely this should also move to the HarnessMessageConverter somehow

Agreed. The challenge is that buildAssistantMessage needs access to inlineToolNames (the set of BPMN element names) to filter tool calls, which is context the converter doesn't currently have. I can either:

Pass inlineToolNames as a parameter to the converter

Keep filtering in the adapter and move only the message building to the converter

WDYT?

nikonovd · 2026-05-27T12:32:09Z

+   * @param toolUseBlock the Harness tool use block
+   * @return Camunda ToolCall
+   */
+  public ToolCall toToolCall(HarnessToolUseBlock toolUseBlock) {


❓ Where is this method being used? I can't see it being used here besides the test class.

You're right it was created anticipating we'd need to convert HarnessToolUseBlock from the response but we don't need it so I'll remove it.

+            AwsBasicCredentials.create(staticAuth.accessKey(), staticAuth.secretKey());
+        builder.credentialsProvider(StaticCredentialsProvider.create(credentials));
+      }
+      case AwsAuthentication.AwsDefaultCredentialsChainAuthentication ignored ->


+        builder.credentialsProvider(StaticCredentialsProvider.create(credentials));
+      }
+      case AwsAuthentication.AwsDefaultCredentialsChainAuthentication ignored ->
+          builder.credentialsProvider(DefaultCredentialsProvider.create());


+      }
+      case AwsAuthentication.AwsDefaultCredentialsChainAuthentication ignored ->
+          builder.credentialsProvider(DefaultCredentialsProvider.create());
+      case AwsAuthentication.AwsApiKeyAuthentication ignored ->


- Rename template to 'Amazon Bedrock AgentCore Managed Agent (Alpha)' - Add [Alpha] prefix to description and 'Amazon' keyword - Update engine version to ^8.10 - Revert log level from INFO to DEBUG - Rename AgentCoreHarnessJobWorker to AgentCoreHarnessSubProcess - Add session timeout documentation comment - Remove unused toToolCall method from HarnessMessageConverter

tshepomaredi added 2 commits May 11, 2026 09:46

github-advanced-security AI found potential problems May 20, 2026

View reviewed changes

test(agentic-ai): add unit tests for AgentCore Harness converters

882d64d

- Add HarnessToolConverterTest (7 tests) - Add HarnessMessageConverterTest (12 tests) - Cover tool definition conversion, message conversion, and error handling

tshepomaredi added the agentic-ai label May 20, 2026

github-advanced-security AI found potential problems May 20, 2026

View reviewed changes

Comment thread ...java/io/camunda/connector/agenticai/aiagent/agentcoreharness/AgentCoreHarnessSubProcess.java Fixed

tshepomaredi marked this pull request as ready for review May 20, 2026 09:56

tshepomaredi requested review from a team as code owners May 20, 2026 09:56

tshepomaredi requested review from maff, mathias-vandaele and nikonovd May 20, 2026 09:56

tshepomaredi self-assigned this May 20, 2026

nikonovd requested changes May 27, 2026

View reviewed changes

tshepomaredi assigned maff, nikonovd and mathieu-stennier May 29, 2026

tshepomaredi requested a review from nikonovd May 29, 2026 10:01

github-advanced-security AI found potential problems May 29, 2026

View reviewed changes

tshepomaredi force-pushed the tshepo-agentcore-harness-adapter branch from 228f99f to aba3ff4 Compare May 29, 2026 11:17

Merge branch 'main' into tshepo-agentcore-harness-adapter

7fe879f

nikonovd removed the request for review from mathias-vandaele June 1, 2026 12:12

maff marked this pull request as draft June 2, 2026 09:19

	"keywords": ["AI", "AI Agent", "AWS", "AgentCore", "Harness", "Bedrock"],
	"keywords": ["AI", "AI Agent", "Amazon", "AWS", "AgentCore", "Harness", "Bedrock"],

Conversation

tshepomaredi commented May 20, 2026

Description

Related issues

Checklist

Uh oh!

CLAassistant commented May 20, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants