Move a Python Agents SDK agent toward production
Use this skill for microsoft_agents.hosting.core, microsoft_agents.hosting.aiohttp, and related SDK packages. Preserve the selected scenario and change only
files that directly run, host, secure, observe, test, deploy or operate the agent.
An external dependency requiring changes outside that scope needs an explicit
scope decision. A question or review request is read-only.
Sources
- Read Prepare an agent for production on Microsoft Learn first, using the Python language selection. Treat the article and its linked shared production controls as normative.
- Read the packaged readiness matrix for control IDs, statuses, applicability, dependencies and evidence rules.
- Inspect
samples/python/production-reference/, or the reference. It demonstrates one bounded Web Chat/App Service/Blob profile. - Inspect the target's actual SDK APIs and dependency baseline. The reference uses stable released packages; verify the installed APIs before changing the user's dependency version or selecting production controls.
Workflow
1. Discover
- Determine whether the user wants a new sample, an assessment, or changes to an existing sample. Do not ask when the request or repository already shows the answer.
- Inspect the candidate Agents SDK sample before asking questions. If multiple samples exist, identify the requested sample or ask which one is in scope.
- Identify channel, cloud, host, inbound audiences and issuers, trusted calling applications, downstream identities, outbound hosts, state, external dependencies, data classes, retention, scale, side effects, long-running work, model calls, retrieval, tools, token-bearing downloaders, attachments, transcripts, and proactive messaging.
- State the deployment boundary. Do not broaden channel, host, identity, data, or application scope without user approval.
2. Assess
- Classify every applicable readiness control as
verified,implemented,missing,not applicable, orblockedusing the readiness matrix. - Cite repository evidence for each
verifiedorimplementedresult. Do not infer production evidence from the presence of source code. - For review or question-only requests, stop at guidance and assessment. Do not modify files.
3. Agree
- Recommend the smallest suitable production profile and defaults.
- Ask only for unresolved decisions that materially affect architecture, cost, identity, channel, data handling, or side effects.
- Present the implementation sequence and identify changes that would extend beyond Agents SDK-related code. Obtain approval for that expansion.
4. Implement
- Add required controls first. Add conditional controls only for capabilities present in the selected scenario.
- Preserve the agent's intended behavior. Do not refactor unrelated code as part of production hardening.
- Add or update tests and operations documentation with each control. Keep secrets and sensitive user content out of source, generated files, logs, errors, prompts, and telemetry.
5. Verify
- Run the repository-supported build, lint, unit, and integration checks that apply to the changed Agents SDK code.
- Exercise failed production configuration, authentication rejection, issuer and audience/lifetime rejection, applicable caller authorization, service URL claim and outbound-host validation, payload rejection, safe errors, probes, graceful shutdown, and telemetry shutdown/redaction.
- Validate state recovery and conflict handling only when state exists. Validate conditional controls only for capabilities in scope.
- Separate local verification from deployed evidence. Do not claim that a control is verified when its required environment was not exercised.
6. Handoff
Report:
- The selected deployment boundary and current maturity state.
- Controls verified with file, test, deployment, or operator evidence.
- Controls implemented but not verified in the required environment.
- Missing or blocked controls and their production impact.
- Controls marked not applicable and why.
- The next deployment, smoke-test, rollback, and operations actions.
Keep this assessment current through follow-up discussion. Answer questions without losing the selected boundary, decisions, statuses, or next phase.
Maturity states
- Code-hardened: applicable source controls exist and local checks pass.
- Deployment-ready: infrastructure, production configuration, smoke tests, rollback, and operator documentation exist.
- Production-verified: the deployed environment has passed authentication, state recovery where applicable, telemetry, alert, traffic-protection, smoke, and rollback checks.
Language invariants
- For aiohttp, use SDK JWT middleware/decorators exactly once on protected routes. Keep health endpoints outside authentication and explicitly protect new custom routes.
- Set
AgentAuthConfiguration.validate_issuer=True, require exact audience and expiration, disable anonymous production access, and select the expected token family, tenant and cloud. Authorize verifiedclaims_identitycaller claims for trusted-app endpoints before adapter processing. - Verify the issuers actually accepted by the installed SDK. Its trusted
first-party Entra issuers can extend the configured issuer list. For the bounded
public-cloud Bot Framework profile, additionally require the verified
claims_identityissuer to equalhttps://api.botframework.comafter JWT validation and before adapter processing. Other profiles need their own issuer policy; do not apply this Bot-only restriction to trusted-app endpoints. - Pass
OutboundHostValidatorexplicitly toCloudAdapter. Confirm downloader support before sharing it; host suffix validation alone does not provide DNS or arbitrary-client protection. - Use the Blob service URL (not container URL) with
BlobStorageConfigand async credentials. The released SDK overwrites state unconditionally; use the reference's ETag wrapper or an equivalent durable concurrency-safe provider when needed. - Preserve async cleanup: drain aiohttp, close clients/credentials, then flush telemetry. Verify published package versions and deployment wheel hashes.
Shared invariants
- Production state that must survive turns/restarts/replicas requires a durable store. Stateless agents do not need a state provider.
- Wildcard connection selection is not authorization. Keep exact audience, verified issuer/caller checks, and outbound policy.
- Validator host entries can allow subdomains. Exact channel middleware and arbitrary-client/network controls remain separate decisions.
- Health endpoints may be anonymous; routes with mutations/downstream access require authentication and applicable authorization.
- Caller-aware edge protection, deployed alerts, recovery and rollback require environment evidence. Do not certify a deployment from sample code.
- Keep secrets, tokens, prompts/messages and unnecessary identity attributes out of errors, logs, generated artifacts and telemetry.