Opt into the Strands (Python) Converse Lambda

Starting in 0.25.0, Pika ships an optional Python converse Lambda built on the Strands Agents SDK as a drop-in replacement for the default TypeScript converse Lambda. This guide walks through when to use it, how to turn it on, and what changes under the hood.

What You'll Accomplish

By the end of this guide, you will:

• Understand how the Strands path differs from the TypeScript converse Lambda
• Enable the opt-in ConverseStrandsConstruct via pika-config.ts
• Swap your chat app's SSM lookup to route traffic through the Strands Lambda
• Configure your local and CI bundling environment (arm64 + QEMU)
• Verify the cutover in CloudWatch

Prerequisites

• Pika 0.25.0 or later
• A working Pika deployment on AWS (both pika service stack and pika-chat stack)
• Docker running locally for CDK Python Lambda bundling
• On x86_64 CI runners: QEMU + buildx available (see CI/CD Bundling)

When to Use the Strands Lambda

The Strands Lambda is a full-parity alternative to the TypeScript converse Lambda. It exists because the default path delegates the agent tool loop entirely to Bedrock via InvokeInlineAgentCommand, which limits what Pika can observe or override. Strands runs the loop itself, giving you:

• Fewer LLM calls — Semantic directives are modeled as Strands Skills instead of a separate Nova Lite selection call. Typical savings of one model invocation per request
• Iteration & budget control — Cap tool-call iterations, apply time budgets, and intercept individual tool calls
• Prompt caching — CacheConfig(strategy="auto") on the Bedrock model (not available through InvokeInlineAgentCommand)
• Native streaming — Real-time trace, heartbeat, and pika-metadata emission via a queue-based streaming loop

Early A/B benchmarks on representative workloads showed meaningful latency, cost, and tool-call reductions versus the TypeScript path. Actual numbers depend heavily on agent shape, tool count, and model mix — treat the Strands path as an opt-in experiment and run your own A/B before cutting production traffic over.

Note

Feature parity, not a fork. The Strands Lambda supports the same 15 features as the TypeScript Lambda: JWT auth, dynamic agent/tool loading from DynamoDB, tag definitions, semantic directives, entity access control, user instructions, knowledge base integration, collaborator / supervisor orchestration (via Strands Swarm), file uploads, user memory (Bedrock AgentCore), real-time streaming, title generation, answer verification, pricing/usage tracking, and session insights.

Architecture

The Strands Lambda is deployed as a framework-managed opt-in construct. When the feature flag is on, the pika service stack instantiates a ConverseStrandsConstruct alongside the TypeScript converse Lambda — both exist simultaneously, each publishing their own SSM parameter:

• /stack/<proj>/<stage>/function/converse_url — TypeScript Lambda (default)
• /stack/<proj>/<stage>/function/converse_strands_url — Python/Strands Lambda

The chat app picks which URL to call by reading one of these SSM parameters in apps/pika-chat/infra/lib/stacks/pika-chat-construct.ts. Swapping between the two is a one-line change in that file.

┌──────────────────┐
│   Chat App       │
│ (pika-chat)      │
└────────┬─────────┘
         │ reads one SSM param at deploy time
         ▼
   /function/converse_url  ◀─── ConverseFn (TypeScript, default)
   /function/converse_strands_url  ◀─── ConverseStrandsFunction (Python, opt-in)

Runtime and Bundling

• Runtime: Python 3.14
• Architecture: Graviton (arm64)
• Response streaming: AWS Lambda Web Adapter layer (LambdaAdapterLayerArm64)
• Container: FastAPI wrapper + queue-based streaming so Swarm's stream_async can coexist with Lambda Web Adapter's event loop

Step 1: Enable the Feature Flag

Edit pika-config.ts at the repository root:

export const pikaConfig: PikaConfig = {
    // ...
    siteFeatures: {
        // ...
        intentRouter: { enabled: true },
        strandsConverse: { enabled: true }  // ← add this
    },
    // ...
};

This tells pika-stack.ts to instantiate the ConverseStrandsConstruct. The TypeScript converse Lambda is always created regardless — the flag only controls whether the Strands Lambda is also deployed.

Step 2: Route Traffic to the Strands Lambda

Edit apps/pika-chat/infra/lib/stacks/pika-chat-construct.ts and change the SSM parameter name that the chat app reads:

// Before (default — routes to TypeScript converse Lambda)
const converseFnUrl = ssm.StringParameter.valueForStringParameter(
    this,
    `/stack/${props.pikaServiceProjNameKebabCase}/${props.stage}/function/converse_url`
);

// After (routes to Strands Python Lambda)
const converseFnUrl = ssm.StringParameter.valueForStringParameter(
    this,
    `/stack/${props.pikaServiceProjNameKebabCase}/${props.stage}/function/converse_strands_url`
);

Tip

You can roll out gradually by only flipping this in your non-production stages first. Run the Strands Lambda alongside production traffic for a week, compare latency and cost in CloudWatch, then cut prod over with confidence.

Step 3: Configure Bundling

Local Development

CDK uses Docker to build the Python Lambda for linux/arm64. On an Apple Silicon Mac, this runs natively. On Intel/AMD machines, Docker Desktop handles emulation transparently.

Verify Docker is running before cdk synth or cdk deploy:

docker info

CI/CD Bundling (arm64 + QEMU)

Standard x86_64 GitHub Actions runners cannot execute arm64 Docker images without emulation. Add QEMU + Buildx to your deploy workflow before the CDK deploy step:

# .github/workflows/deploy.yml (or equivalent)
- name: Set up QEMU
  uses: docker/setup-qemu-action@v3

- name: Set up Docker Buildx
  uses: docker/setup-buildx-action@v3

- name: CDK deploy
  run: pnpm --filter @pika/service cdk:deploy

Bundling the Python Lambda under QEMU takes roughly 60–120 seconds per build. If that becomes a bottleneck, provision native arm64 runners instead of emulating — both approaches work; this guide doesn't prescribe one.

Caution

Skipping QEMU on x86_64 CI produces an exec format error or — more insidiously — a Lambda that fails at INIT with ModuleNotFoundError: No module named 'pydantic_core._pydantic_core', because pip pulls x86_64 manylinux wheels whose compiled .so extensions can't load on Graviton. Always confirm bundling completed on an arm64-compatible host.

Step 4: Local Development Tooling

The Strands Lambda ships with helper scripts in services/pika/package.json:

Command	What it does
pnpm --filter @pika/service strands:setup	Creates a .venv, installs Python + dev dependencies
pnpm --filter @pika/service strands:test	Runs the Python pytest suite (contract + unit tests)
pnpm --filter @pika/service strands:invoke	Invokes the handler locally against real DDB + Bedrock
pnpm --filter @pika/service strands:validate	Full loop: setup, test, CDK synth

Run strands:setup once per clone, then strands:test on each change. strands:invoke requires AWS credentials and an agent definition already seeded in DynamoDB.

Step 5: Deploy and Verify

Deploy the service stack first (creates the Lambda + SSM params), then the chat stack (reads the param you flipped):

pnpm --filter @pika/service cdk:deploy
pnpm --filter @pika/chat cdk:deploy

After deploy, verify:

1. Lambda exists and INIT succeeds — check /aws/lambda/<project>-<stage>-ConverseStrandsFunction... in CloudWatch Logs. You should see Uvicorn running on http://0.0.0.0:8080. If you see ModuleNotFoundError, the bundling platform mismatched the runtime arch (see the caution box above).

2. Traffic is flowing to Strands — send a chat message and tail the same log group. You should see the request handler entry and subsequent streaming traces. If you see no traffic, confirm the chat stack was deployed after flipping the SSM path.

3. SSM parameters are correct:

aws ssm get-parameter --name /stack/<proj>/<stage>/function/converse_strands_url
aws ssm get-parameter --name /stack/<proj>/<stage>/function/converse_arn

Rolling Back

The TypeScript Lambda stays deployed while the Strands flag is on. To roll back:

1. Change the SSM lookup in pika-chat-construct.ts back to converse_url
2. Redeploy the chat stack

Traffic cuts back to the TypeScript path immediately. You can leave strandsConverse.enabled = true if you want to keep the Strands Lambda warm for experimentation — or flip it off to stop paying for an idle function.

Migrating from an Existing Strands Customization

If you previously added a Strands-style Lambda via services/pika/lib/stacks/custom-stack-defs.ts (customer override), migrate to the framework construct after upgrading to 0.25.0:

1. Delete your custom createStrandsConverseLambda() implementation and any supporting constants from custom-stack-defs.ts. Restore the file to the pika scaffold defaults.
2. Flip pikaConfig.siteFeatures.strandsConverse.enabled to true.
3. Deploy.

CloudFormation will replace the old resources (deleting the customer-side logical IDs and creating new framework-side ones). The Function URL will change, but the /function/converse_strands_url SSM parameter rewrites automatically — downstream consumers picking it up through SSM will cut over on their next deploy. Expect a brief reachability gap at cutover; coordinate an off-hours deploy if needed.

What's Next

• Compare against the TypeScript path in a non-prod stage. Deploy both Lambdas side-by-side, then alternate the SSM lookup in pika-chat-construct.ts to A/B them. CloudWatch metrics (Duration, Invocations, Throttles) on the two function names give you the comparison data.
• Review the Strands Agents SDK documentation. Customization the framework doesn't expose — tool loops, reasoning traces, custom callbacks — lives at the SDK layer. Start with strands-agents/sdk-python.
• Understand the directive pathway. The Strands path prefers Strands Skills for directive selection (no LLM call) and falls back to the same Nova Lite filtering the TS Lambda uses when Skills aren't viable. If you use semantic directives heavily, the Instruction Augmentation guide still applies — directive authoring and DDB storage are unchanged.
• Keep customizations out of handler.py. Constants like MAX_ITERATIONS and LAMBDA_TIMEOUT_BUFFER_SECONDS live in framework code that pika sync will overwrite. If you need to tune them, open an issue on the Pika repo so the knob becomes configuration rather than a fork of the Lambda source.