Build

Every module in Learn is backed by a production lab in labs/. A lab isn’t a code snippet — it’s a service with its own test suite, CI pipeline, and production readiness notes, built to the bar described below.

What makes a lab “production”

A lab is graded against the same rubric a real system design review would use:

• Architecture that survives contact — clear module boundaries, explicit interfaces between them, and a documented request flow.
• Failure handling, not just the happy path — timeouts, retries, backpressure, graceful shutdown, and draining are implemented, not described.
• Tests that would catch a regression — unit tests for logic, integration tests for the pieces that talk to Redis/Kafka/etc., and load tests for anything claiming a throughput number.
• Observability built in — structured logs, metrics, and traces wired up from the start, not bolted on before a demo.
• CI that actually gates merges — lint, type-check, and test, running in .github/workflows/ against every change to the lab.
• A production readiness doc — what’s load-tested, what’s known-broken, and what you’d need to change before pointing real traffic at it.

Labs

⚙Hands-on Lab

An async API gateway in front of multiple LLM providers: JWT-based tenant identity, Redis-backed distributed rate limiting, health-aware provider fallback, graceful connection draining, and OpenTelemetry tracing. Read the lab documentation →

labs/async-ai-gatewayproduction-ready

⚙Hands-on Lab

Lease-based task delivery with visibility timeouts and fencing tokens, idempotent submission, checkpointed resumption, and dead-lettering counted by delivery attempt rather than explicit failure. Read the lab documentation →

labs/durable-agent-task-engineproduction-shaped

⚙Hands-on Lab

Latency-bounded request batching that races a size trigger against a timeout, canary version routing with independent per-version metrics, and a tail-latency benchmark harness. Read the lab documentation →

labs/dynamic-batching-inferenceproduction-shaped

⚙Hands-on Lab

Structure-aware chunking, BM25 and vector retrieval fused by rank rather than score, a reranking stage, groundedness checking, and an evaluation harness that measures retrieval quality instead of assuming it. Read the lab documentation →

labs/hybrid-retrievalproduction-shaped

⚙Hands-on Lab

What it costs to checkpoint a LangGraph state that grows every step: a quadratic write curve against DeltaChannel’s linear one, the resume cost that buys it, and a real deadlock in the pinned release. Read the lab documentation →

labs/langgraph-checkpoint-costproduction-shaped

⚙Hands-on Lab

A real MCP server on the official SDK speaking the 2026-07-28 protocol: transport-level tenant authentication, per-tenant discovery, refusals indistinguishable from “unknown tool”, and statelessness proven by test. Read the lab documentation →

labs/multi-tenant-mcp-serverproduction-shaped

⚙Hands-on Lab

The enforcement layer in front of agent tool execution: capability scoping, schema validation, per-tool rate limiting, human approval gates with a bounded wait, and an audit log that records denials as carefully as successes. Read the lab documentation →

labs/policy-gated-tool-runtimeproduction-shaped

⚙Hands-on Lab

How much of an RTO a failover trigger spends deciding to fail over: a window tied to the RTO guarantees missing it, silence reads as health, and the eager alternative fails over on blips. Read the lab documentation →

labs/region-failover-budgetproduction-shaped

⚙Hands-on Lab

Error-budget tracking with multiwindow burn-rate alerting, an autoscaling controller whose cooldown a fast burn deliberately overrides, and incident runbook execution that escalates rather than retrying — all reading one shared severity signal. Read the lab documentation →

labs/slo-driven-ai-operationsproduction-shaped

Lab	Status
Async AI Gateway	Production-ready
SLO-Driven AI Operations	Production-shaped
Durable Agent Task Engine	Production-shaped
Policy-Gated Tool Runtime	Production-shaped
Hybrid Retrieval and Evaluation	Production-shaped
Dynamic Batching Inference	Production-shaped
Multi-Tenant MCP Server	Production-shaped
LangGraph Checkpoint Cost	Production-shaped
Agent Identity Broker	Production-shaped
Model Router	Production-shaped
Semantic Cache	Production-shaped
Evaluation Platform	Production-shaped
Region Failover Budget	Production-shaped

What the statuses mean

Production-ready means you could point real traffic at it: container, deployment manifest, load profile, and a production-readiness document. Production-shaped means the architecture, tests, and failure handling are real but at least one production dependency is deliberately simulated — each lab’s README names exactly which. Passing tests alone never earns production-ready.