Skip to content

AgentCore Support (3/3) - Integrate AgentCore into SKILL.md, README and concepts.md - #27

Open
justinschoeff wants to merge 3 commits into
agentcore/sdk-guidesfrom
agentcore/skill-integration
Open

justinschoeff wants to merge 3 commits into
agentcore/sdk-guidesfrom
agentcore/skill-integration

Conversation

@justinschoeff

@justinschoeff justinschoeff commented Oct 7, 2026 •

Copy link
Copy Markdown

AgentCore Support (3/3). Stacked on #26 (2/3), which is stacked on #25 (1/3) → main. The stack is rebased onto #31 and #30.

This PR connects references/aws-agentcore/ to the skill's entry points, so agents send AWS users to AgentCore when it fits.

What changed in this update

Changes

SKILL.md

  • Frontmatter description: adds Temporal-specific AgentCore triggers. No version: field.
  • Support table: AgentCore Runtime is listed as Supported — Pre-release, with its own SDK table (Python, TypeScript).
  • Step 1:
    • A GCP Namespace still maps to Cloud Run.
    • An AWS Namespace gets a structured AgentCore-vs-Lambda question, with criteria from Choose AgentCore or Lambda.
    • On Cloud, the user confirms the Namespace has Pre-release access. If it doesn't, the skill points them to support or their account team, with self-hosting as the alternative.
  • Step 2: an AgentCore approval list that names every created resource:
    • the CDK stack;
    • for TypeScript, the Container build's ECR repository, KMS key, CodeBuild project and build Lambda;
    • Transaction Search.
  • "AgentCore: steps 3–8":
    • CodeZip for Python and Container for TypeScript, with no local Docker;
    • the API-key hand-off;
    • waiting on endpoint READY from the AWS API;
    • warm and optional scale-from-zero verification from the endpoint's log group;
    • inventory and teardown.
  • Routing rows, WCI diagnostics pointers and troubleshooting entries for AgentCore.
  • Step 1 notes that tcld namespace get does not report AgentCore Pre-release access, so the skill asks the user to confirm it.

README.md: AgentCore is listed in one support-table row as Supported — Pre-release (Python, TypeScript), with access notes, prerequisites, stages and repository-guide rows.

references/concepts.md: a provider-neutral pointer to aws-agentcore/constraints.md and an AgentCore entry in the provider table.

End-to-end validation

End-to-end deploys (following the skill from Step 1)

Both runs used Temporal Cloud Namespace quickstart-schoeff.a2dd6, AWS account 429901336738 in us-east-1, Temporal CLI 1.8.3 and the AgentCore CLI 0.28.1. In both runs, Step 1 asked the AgentCore-vs-Lambda question as a structured choice, and AgentCore was chosen. The user confirmed Pre-release access. The Step 2 approval list was approved before anything was created. The API key went into Secrets Manager without being printed.

Python TypeScript
Worker samples-python bedrock_agentcore/strands_agent (9d5c46b), with the sdk-python.md changes: key from TEMPORAL_API_KEY_SECRET_ARN, required env vars minimal Worker built only from sdk-typescript.md (greetingWorkflow + greet)
Build AgentCore CLI, CodeZip AgentCore CLI, Container (CodeBuild → ECR)
Invocation role assets/temporal-cloud-serverless-worker-agentcore-role.yaml, stack schoeff-skillval-invoke (shared; second Runtime added with update-stack) same stack
Worker Deployment schoeff-skillval-py, build v3, endpoint temporal_v3 schoeff-skillval-ts, build v1, endpoint temporal
Workflow ID schoeff-skillval-py-e2e-v3-1792784267 schoeff-skillval-ts-e2e-1792773674
Status Completed, Pinned to schoeff-skillval-py:v3. 3 Activities (model, code interpreter, model): "The 30th Fibonacci number is 514229, and it is not divisible by 7." Completed, Pinned to schoeff-skillval-ts:v1: "Hello, schoeff-skillval, from an AgentCore TypeScript Worker"
Worker log polling schoeff-skillval-py-tq as schoeff-skillval-py/v3 polling schoeff-skillval-ts-tq as schoeff-skillval-ts/v1

Python took three builds, and each build got its own named endpoint, as versioning.md prescribes:

  • v1: the sample as-is. The Worker registered and ran, but the sample's default Anthropic model was blocked: the account hadn't submitted the Anthropic use-case form.
  • v2: set an Amazon Nova model through StrandsPlugin(models=…). Strands then requires TemporalAgent(model="bedrock"), which is a Workflow code change.
  • v3: completed.

The model swap is agent code, which this skill links to rather than covers. Everything the run created was torn down afterwards.

Fixes made from the runs

Folded into #25, #26 and this PR (each PR is now one squashed commit):

  1. TypeScript must use the Container build. The CLI's Node CodeZip packager esbuild-bundles main.ts alone and ships no node_modules, so agentcore deploy failed at synth. The packager can't load @temporalio/worker's native swc and core-bridge modules, and the CLI can't take a prebuilt zip.
  2. The TypeScript image needs two changes. It must install ca-certificates, because the Worker failed with TransportError … NativeCertsNotFound on node:22-slim. It must also run compiled dist/main.js, because the CLI template's npx tsx main.ts breaks workflowsPath: './workflows.js'.
  3. Read Worker logs from the named endpoint's log group (/aws/bedrock-agentcore/runtimes/<RUNTIME_ID>-<ENDPOINT>). agentcore logs --runtime returned nothing.
  4. Read endpoint READY from aws bedrock-agentcore-control get-agent-runtime-endpoint. agentcore status only reports deployed.
  5. Apply the naming prefix to the AgentCore project name. It names the CDK stack and prefixes the Runtime ID. Fill in aws-targets.json, which agentcore create leaves as [].
  6. Resolve additionalPolicies relative to codeLocation, not the project root. The TypeScript synth failed with ENOENT.
  7. Approval list and inventory now include the CDK stack, Container resources, log groups and account-wide CloudWatch Transaction Search, which agentcore deploy turned on.
  8. iam.md shows how to add a Runtime to an existing invocation stack (escaped comma in AgentRuntimeARNs).
  9. sdk-python.md notes that the sample needs Bedrock model access.
  10. Teardown deletes leftover log groups. Deleting a Runtime and its stack left one /aws/bedrock-agentcore/runtimes/<RUNTIME_ID>-<ENDPOINT> group per endpoint, including DEFAULT. A Container build also left its CodeBuild and build-trigger Lambda log groups.

@justinschoeff
justinschoeff requested a review from a team as a code owner October 7, 2026 19:12

@starfleeth starfleeth left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Could you remove the local/scripted validation narration and the promise that future manual-deployment results will be added later? Please either report completed end-to-end validation with stable evidence or leave an explicit unchecked readiness item and avoid presenting Python and TypeScript integration as fully ready until it is complete.

Comment thread SKILL.md Outdated

7. **Verify.** Start a Workflow, confirm its history progresses, and confirm the AgentCore logs show the Worker starting, processing Tasks, and draining. If it does not progress, → `references/aws-agentcore/diagnostics.md`.

8. **Hand back the inventory first; offer teardown as the closing note.** Include the items in `references/aws-agentcore/setup.md` under "Resource inventory". The skill's sources do not cover AgentCore teardown commands, so ask the user how they want AgentCore resources removed before deleting any.

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Could you remove “the skill’s sources do not cover…” and retain the direct safety requirement? For example: “Do not delete AgentCore resources automatically. Ask the user which supported removal method they want before removing anything.”

@justinschoeff
justinschoeff force-pushed the agentcore/skill-integration branch from 2ea9d62 to 74d6dee Compare October 26, 2026 19:20
@justinschoeff
justinschoeff force-pushed the agentcore/skill-integration branch from 74d6dee to 734282b Compare October 26, 2026 19:28
@justinschoeff
justinschoeff force-pushed the agentcore/skill-integration branch from 734282b to e35b3f6 Compare October 26, 2026 19:35
@justinschoeff
justinschoeff force-pushed the agentcore/skill-integration branch from e35b3f6 to cc6bbfd Compare November 6, 2026 20:55
@justinschoeff justinschoeff changed the title AgentCore Support (3/3) - Add AgentCore to SKILL.md and README AgentCore Support (3/3) - Integrate AgentCore into SKILL.md, README and concepts.md Nov 6, 2026
Includes the follow-up fixes: Temporal-specific triggers, bounded
waits, Container build for TypeScript, the AgentCore-or-Lambda
justification, one Pre-release README row and the concepts provider
table entry.
docs.temporal.io/cloud/namespaces#access-namespaces says that on a
Namespace with both API keys and mTLS enabled, API keys cannot
authenticate through the Namespace endpoint, so API-key clients must
use the regional endpoint. tcld reports such a Namespace as
.spec.authMethod ApiKeyOrMtls.

The old text said both endpoints accept API keys, which is wrong for
these Namespaces.
docs.temporal.io/cloud/namespaces#access-namespaces says that on a
Namespace with both API keys and mTLS enabled, API keys cannot
authenticate through the Namespace endpoint, so API-key clients must
use the regional endpoint. tcld reports such a Namespace as
.spec.authMethod ApiKeyOrMtls.

The Lambda examples hard-code the tmprl.cloud Namespace endpoint, which
fails for API-key clients on these Namespaces.
@justinschoeff
justinschoeff force-pushed the agentcore/skill-integration branch from cc6bbfd to 48d2fb3 Compare November 9, 2026 16:13
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants