Lab Production Readiness Checklist
The labs are teaching implementations. This checklist defines what must be added before a lab pattern becomes production work.
Use this after completing a lab. The goal is to identify the next engineering boundary: persistence, authorization, retries, idempotency, observability, eval gates, deployment, and rollback. For the full production path, continue with Deployment Walkthrough, Templates and Worksheets, and the 10/10 Production Gate.
Download the reusable production review artifact: 10/10 production gate scorecard.
Download the lab-specific worksheet: lab production readiness worksheet.
Universal Production Gate
Every lab needs these controls before real users, real data, or real side effects:
| Gate | Required Evidence |
|---|---|
| State ownership | State schema, owner, persistence strategy, migration plan, and replay story. |
| Auth and permissions | Actor identity, tenant/resource scope, tool permissions, approval rules, and audit records. |
| Tool safety | Input schemas, side-effect class, idempotency key, timeout, retry policy, and error contract. |
| Memory safety | Retention, deletion, redaction, consent, correction path, and write policy. |
| Observability | Trace IDs, model/tool events, policy decisions, costs, latency, stop reasons, and redaction. |
| Eval gates | Golden tasks, negative cases, trajectory checks, safety checks, and release-blocking thresholds. |
| Deployment | Runtime owner, environment config, secrets, scaling limits, rollback, kill switch, and incident path. |
| Human control | Approval UI/API, escalation rules, cancellation, reviewer identity, and expiry. |
sequenceDiagram
participant Lab
participant Reviewer
participant Gate as Production gate
participant Runtime
participant Operators
Lab->>Reviewer: Baseline command, output, source files
Reviewer->>Gate: Check state, auth, tools, memory, traces, evals
alt Gate incomplete
Gate-->>Reviewer: Keep as lab evidence
else Gate complete
Reviewer->>Runtime: Name added authority and rollout stage
alt No real users, data, or side effects
Runtime-->>Reviewer: Prototype with synthetic data
else Real authority involved
Runtime->>Gate: ADR, policy boundary, trace contract, eval fixtures
Gate->>Operators: Dashboard, incident path, rollback, kill switch
Operators-->>Reviewer: Can disable, replay, and explain
end
end
Use this flow before changing access, data, or authority. A lab can move forward only when the next stage has reviewable evidence, not because the example command ran once.
Lab Review Questions
Use these questions before a lab-derived system becomes a product slice:
| Question | Required Evidence |
|---|---|
| What did the lab prove? | Baseline command, expected output, source files, and success signal. |
| What did the lab not prove? | Missing persistence, policy, observability, evals, deployment, rollback, or human control. |
| What authority will production add? | Read data, write tools, memory, messages, approvals, money movement, or workflow execution. |
| What must fail closed? | Missing config, missing policy context, unsafe tool input, stale evidence, and failed evals. |
| What is the next production artifact? | ADR, trace contract, eval fixture, runbook, checklist, dashboard, or rollback plan. |
The lab is a learning artifact until these answers exist in reviewable engineering documents.
Promotion Ladder
Use this ladder to decide what the lab output is ready for.
| Stage | Allowed Use | Required Evidence | Stop Condition |
|---|---|---|---|
| Lab evidence | Learning, comparison, and design discussion. | Baseline command, success signal, inspected source files, and one known failure path. | Do not connect real users, private data, credentials, or side effects. |
| Prototype | Internal demo with synthetic data. | Lab evidence plus basic schemas, deterministic test, and named production gaps. | Stop if the demo requires real credentials or broad tool access. |
| Product slice | Limited internal workflow with controlled data. | ADR, policy boundary, trace contract, eval fixtures, owner, and rollback note. | Stop if policy, trace, or rollback is missing. |
| Pilot | Small real-user or real-data rollout. | Production gate scorecard, approval rules, incident path, dashboard, and release gate. | Stop or roll back on failed blocking evals, missing trace spans, or unsafe side effects. |
| Production candidate | Controlled production release. | Deployment walkthrough evidence, runbook, kill switch, rollback path, and current eval report. | Stop if operators cannot disable, replay, or explain the run. |
Do not skip stages because a framework example runs successfully. A passing command proves execution; it does not prove authority, observability, recovery, or release safety.
Per-Lab Readiness Matrix
| Lab | Production Additions |
|---|---|
| Lab 01 - Tool-Using Agent | Tool schemas, side-effect labels, permission checks, idempotency, timeout/retry, and audit records. |
| Lab 02 - Agent Loop and Planning | Durable state, plan versioning, step retry policy, cancellation, partial failure handling, and plan evals. |
| Lab 03 - Agentic RAG | Source ACLs, freshness, citation checks, retrieval evals, prompt-injection filtering, and evidence retention rules. |
| Lab 04 - A2A Communication | Agent identity, signed envelopes, correlation IDs, cancellation semantics, schema versioning, and replay logs. |
| Lab 05 - Multi-Agent Supervisor | Worker contracts, merge policy, per-worker traces, disagreement handling, final acceptance owner, and cost caps. |
| Lab 06 - Observability and Evals | Trace storage, redaction, eval datasets, release thresholds, incident-to-eval workflow, and dashboard ownership. |
| Lab 07 - Mastra Runtime Packaging | Deployment config, tool/memory policy, workflow retries, eval integration, trace export, and framework upgrade plan. |
| Lab 08 - CrewAI Flows and Crews | Flow checkpoints, role permissions, task schemas, crew output validators, human escalation, and flow acceptance evals. |
| Lab 09 - Minimal Agent Loop | Decision validation, tool registry, stop policy, state persistence, timeout/cancellation, and trace events. |
| Lab 10 - Tool Registry and Policy Gate | Policy context, approval records, schema validation, side-effect isolation, idempotency, and denial analytics. |
| Lab 11 - Context, Memory, Trace, and Evals | Memory governance, trace redaction, eval fixture versioning, context audit, and incident replay. |
| Lab 12 - LangGraph State Graph | Durable checkpointer, thread IDs, interrupt payloads, node idempotency, state migrations, and resume tests. |
| Lab 13 - AutoGen Transcript Evals | Message schemas, team termination, transcript redaction, role permissions, transcript replay, and team-level eval gates. |
Framework-Specific Deployment Questions
| Framework Shape | Production Questions |
|---|---|
| LangGraph-style | Where is the checkpointer stored? How are thread IDs assigned? Which nodes can cause side effects? What state migrations are supported? |
| AutoGen-style | Who owns the transcript? Which messages are durable? How does termination work? How are agent tools permissioned? |
| Mastra-style | Which runtime features are framework-owned? How are workflows deployed? How are tools, memory, traces, and evals exported? |
| CrewAI-style | What does the flow own? What does the crew own? How are role outputs validated? How are crew failures escalated? |
| Mini-runtime | Which production controls are you willing to build and operate yourself? Which should move into an existing workflow platform? |
Release Gate Template
Before shipping a lab-derived system, require:
state: persisted or explicitly stateless
policy: enforced before side effects
tools: typed, scoped, idempotent, timed out
memory: governed by retention/deletion rules
trace: redacted and replayable
evals: include happy path, negative path, and trajectory checks
deployment: rollback and kill switch documented
owner: team and escalation path assigned
If any line is unknown, the system is still a demo.