Timeline
17-18 hours
EXPERT-tier · PRO unlocks Modules 01-03AI & vectors trackP13
Build the
agent execution layer
of modern data platforms
Ship a production agentic pipeline with a LangGraph supervisor + 4 worker agents (Ingestion / Quality / Transform / Loading), Redis checkpointing for time-travel replay, HITL via interrupt_before with Slack actionable approval, ToolCallGuard budget enforcement, RestrictedPython sandbox, and 5 committed ADRs. Modules 01-03 unlock with PRO; the platform unlocks with EXPERT.
The agent-platform system-design portfolio piece for staff AI roles — 5 committed ADRs (one Deprecated documenting a real RBAC reversal), a working LangGraph supervisor-worker pipeline, and a cost model that defends the cascade vs Sonnet-only baseline.
By the end you will have wired
- LangGraph supervisor + 4 worker agents with conditional-edge routing and Redis checkpointing
- RBAC-aware ToolRegistry with per-agent scoped views (post-ADR-005 reversal)
- Production observability: Prometheus + OpenTelemetry + LangSmith + per-agent cost attribution
- HITL approval flow via interrupt_before + Slack actionable buttons (Approve / Deny / Escalate)
- FailureDetector + ToolCallGuard + ContextWindowManager + 24h TimeTravel checkpoint replay
- 5 ADRs (one Deprecated) + cost-model CSV + RestrictedPython sandbox + GitHub PR bot
PREREQ · SENIOR+Built for engineers shipping agents in production. Comfortable with Python services, async / asyncio, at least one of: LangGraph or equivalent agent framework, FastAPI, or production observability. Not a “what is an agent” course.
agentic_pipeline.platform · 6 modules · supervisor + 4 workers armed · LangGraph + Redis checkpoint
interrupt_before ✓Inputs
Agents
Memory
Surfaces
postgresQueryDatabaseTool · pooled
rest_apiAPIClientTool · retry + rate-limit
s3 / filesFileProcessorTool · chunked read
kafkastream consumer · idempotent
Tool registry — see ADR-005 (Deprecated)
SupervisorLangGraph routing · weighted conf
IngestionWorkerRBAC-scoped tool view
QualityWorkervalidation + anomaly
TransformWorkerschema-evolution-safe
LoadingWorkeridempotent writes
Supervisor-worker topology — see ADR-003
RedisSavercheckpoint · 24h TTL
SemanticMemoryembedding + similarity
TimeTravelreplay from any checkpoint
AuditLogtool-call + approval trail
Two-tier persistence — see ADR-002
interrupt_beforeHITL gate · Slack approve
LangSmith traceper-agent + per-tool
Cost trackerUSD per run · per agent
GitHub PR botPyGithub · auto-PR
HITL approval flow — see ADR-004
# Judge cascade — 35% cost cut
Haiku handles 70% of worker LLM calls (USD 0.80/M in)
Sonnet only on supervisor routing + escalation paths
Cascade saves ~USD 75/mo at 5k runs/mo
→ ~USD 0.036 per run at optimized load
# HITL gate — interrupt_before
Destructive writes pause the LangGraph mid-run
State persists in Redis (24h TTL) — no re-execution on resume
Slack actionable buttons (Approve / Deny / Escalate)
→ graph.aupdate_state() resumes from exact checkpoint
4 + 1
workers + 1 supervisor
5 ADRs
committed in starter kit
−35%
cost vs Sonnet-only baseline
Curriculum · 6 modules · 17-18 hours · 3 phases
Modules 01-03 unlock with PRO. The full platform with EXPERT.
Modules 01-03 (~10h) ship a complete working multi-agent pipeline — foundation, 4 production tools, and a LangGraph supervisor + worker system with Redis checkpointing. Included with PRO. Modules 04-06 (~7.5h additional) layer on production observability, hardening (HITL + failure detection + time-travel), and a multi-tenant platform-design capstone. Unlock with EXPERT.
P13 · 6 modules · 17-18 hours · 60+ lessons
Free preview EXPERT required
M01
⊘Agent Foundation
Project scaffolding with LangGraph, base agent class with logging hooks, BaseDataTool abstraction, PipelineState TypedDict, and a Pydantic Settings configuration layer. The honest baseline before any working agent exists.
M02
⊘Tool Implementation
Four production tools — QueryDatabaseTool (connection pooling + SQL safety), APIClientTool (retry + rate-limit + idempotency), FileProcessorTool (chunked read + format detect), ValidationTool (quality rules) — wired through an RBAC-aware registry per ADR-005's reversal.
M03
⊘Multi-Agent Pipeline
LangGraph StateGraph with supervisor + 4 workers, conditional-edge routing, semantic memory store, RedisSaver checkpointing, RBAC scoping at the registry, and end-to-end execution. The PRO finish line: a working agentic pipeline.
M04
⊘Production Ready
Prometheus metrics + OpenTelemetry tracing + LangSmith integration + FastAPI + per-agent cost tracking + Docker + Kubernetes manifests + task queue for horizontal scaling + agent evaluation framework.
M05
⊘Harden the Pipeline
HITL via LangGraph interrupt_before + Slack actionable buttons, FailureDetector (loop + cascade), ToolCallGuard (budget enforcement), ContextWindowManager (rolling window + summarization), TimeTravel (24h Redis checkpoint replay), system-prompt tool-use rules.
M06
⊘Design and Extend an AI Data Platform
Unstructured extraction (pdfplumber + BeautifulSoup), dynamic schema inference (Pandas + Pydantic), RestrictedPython sandbox, multi-tenant TenantContext + RBAC, feedback-driven improvement loop, GitHub PR bot (PyGithub), SLA YAML, and a capstone platform-design exercise.
Modules 01-03 with PRO ($29/mo) · Modules 04-06 with EXPERT ($79/mo)
See plans →Backed by curriculum
Agentic Workflows
Open curriculum
iThis curriculum is the foundation for the project — it’s not a sales add-on. EXPERT subscribers get full access to all modules.
The build, in 3 phases
Foundation. Production. Platform.
Each phase ends with a tagged release, a passing integration test, and a passing red-team failure-injection run. No ambiguity about where you are.
01~10h
Foundation (Modules 01-03)
Working multi-agent pipeline running locally. LangGraph supervisor + 4 worker agents + 4 tools + Redis checkpointing, end-to-end execution on the seeded sample data.
- ✓RBAC-aware ToolRegistry with 4 tools (post-ADR-005 reversal)
- ✓LangGraph supervisor + workers with conditional-edge routing
- ✓RedisSaver checkpointing + recovery on restart
02~2h
Production (Module 04)
Observable, deployable, costable. Prometheus + OTel + LangSmith tracing, per-agent cost attribution, FastAPI service, Docker + Kubernetes manifests, agent eval framework.
- ✓Prometheus + OpenTelemetry + LangSmith integration
- ✓Per-agent + per-tool cost attribution
- ✓Docker + Kubernetes deploy + agent eval suite
03~5.5h
Platform (Modules 05-06)
Hardening + multi-tenant platform design. HITL approval flow, failure detection + recovery, time-travel replay, multi-tenant TenantContext, RestrictedPython sandbox, capstone design.
- ✓HITL via interrupt_before + Slack actionable buttons (ADR-004)
- ✓FailureDetector + ToolCallGuard + TimeTravel replay
- ✓Multi-tenant + RestrictedPython sandbox + GitHub PR bot
Project setup · 10 minutes
One command. Local LangGraph + Postgres + Redis + LangSmith.
What lives in the repo
You get the real agent platform on day one — LangGraph supervisor + 4 worker agents, RBAC-aware ToolRegistry with 4 tools, RedisSaver checkpointing, FastAPI gateway, Prometheus + OpenTelemetry + LangSmith instrumentation, Docker + Kubernetes manifests, plus the M05 hardening (HITL / FailureDetector / TimeTravel) and M06 platform features.
- src/agents/ — supervisor + 4 workers + HITL + FailureDetector + ContextManager
- src/tools/ — 4 production tools + RBAC-aware registry (post-ADR-005)
- src/orchestration/ + src/memory/ — LangGraph StateGraph + RedisSaver + semantic memory
- src/observability/ — Prometheus + OTel + LangSmith + cost-tracking + Slack alerts
- src/api/ + src/scaling/ + Dockerfile + k8s/ — FastAPI + task queue + container + Kubernetes manifests
- docs/adr/ + docs/cost-model/ — 5 committed ADRs (one Deprecated) + the runnable cost-model CSV
Download · Starter Kit · 91 files · 104 KB
Agentic Data Pipeline Starter Kit
Pre-built LangGraph supervisor + 4 worker agents + RBAC tool registry + Redis checkpointing + FastAPI + Docker + Kubernetes. Now bundled: 5 ADR markdown files (docs/adr/) and the runnable cost-model CSV (docs/cost-model/) — unzip and read them straight from the repo.
~/projects/agentic-data-pipeline — zsh
1. Unzip and start the platform
$ unzip agentic-data-pipeline-starter.zip
$ cd agentic-data-pipeline-starter && cp .env.example .env
$ docker compose up -d
2. Run the demo end-to-end
$ export ANTHROPIC_API_KEY=...
$ python scripts/demo.py \
$ --source postgres --checkpoint redis --trace langsmith
3. Open ADR-001 + the cost model
$ less docs/adr/001-langgraph-vs-crewai-vs-custom.md
$ open docs/cost-model/agentic-data-pipeline-cost-model.csv
4. Run the failure-injection suite
$ pytest tests/test_failure_detector.py tests/test_time_travel.py -v
4 + 1
workers + supervisor
91
files in starter kit
5
ADRs (incl. 1 Deprecated)
30
sample run traces
Production hardening
The same agent demo — but built for the auditable case.
Most agent tutorials show you a single LLM call in a loop. This shows what changes when the supervisor decides for 4 workers, every tool call hits an audit log, the cost model is defensible to a CFO, and a compliance reviewer asks which agent decided that.
Notebook agent demoWhat most teams ship
×
Orchestration
While-loop calling an LLM
×
Tool access
Every agent can call every tool
×
Failure mode
Loop until token budget; pray
×
HITL
Manual code edit, redeploy
×
Cost
Whatever the bill says next month
×
Replay
Re-run from scratch on every retry
Your agent platformModules 04–06
✓
Orchestration
LangGraph StateGraph + supervisor-worker conditional edges (ADR-001 + ADR-003)✓
Tool access
ScopedToolView per agent role (RBAC at registry, post-ADR-005 reversal)✓
Failure mode
FailureDetector + ToolCallGuard + budget cap; cascade detection in M05✓
HITL
interrupt_before + Slack approve / deny / escalate (ADR-004)✓
Cost
Per-agent token attribution + judge cascade −35%; CSV in
docs/cost-model/✓
Replay
TimeTravel reads any checkpoint in last 24h via RedisSaver (ADR-002)EXPERT-only · architecture decision records
Write the ADRs staff engineers actually get judged on.
Five ADRs ship inside the starter-kit zip at docs/adr/, one per major decision in the build, including a real Deprecated ADR documenting the v0 ToolRegistry → RBAC reversal after a real Quality-agent overreach incident. The kind of doc that travels with you to your next role. Preview ADR-001 →
ADR-001Accepted
LangGraph chosen over CrewAI / AutoGen / custom orchestrator
Context
HITL + time-travel replay + typed state + streaming all need first-class primitives — pipeline / mesh / chain alternatives don't fit
Decision
Adopt LangGraph as the orchestrator. Workers wrap a Protocol (
WorkerAgent) — orchestrator-replaceableTradeoff
Heavy LangChain dependency surface (~200 transitive deps); minor versions occasionally break Saver protocol
Reversal
Custom orchestrator swap is ~3 engineer-weeks; Protocol shape keeps workers portable
ADR-002Accepted
Redis for orchestrator checkpoints; Postgres only for business data
Context
PostgresSaver caused row-lock contention with M04 dashboards (p95=4.2s → 0.8s after split) + LangGraph version bumps required Postgres migrations
Decision
Two-tier persistence —
RedisSaver for checkpoints (24h TTL), Postgres for business dataTradeoff
Two stores to operate (~$35/mo extra) and 24h replay window vs indefinite
Reversal
Postgres-only re-introduction is ~3 engineer-days when run rate < 500/day
ADR-003Accepted
Hierarchical supervisor-worker topology, not peer-to-peer agents
Context
Per-agent cost attribution + deterministic replay + linear LLM cost scaling all need a single decision-maker
Decision
Supervisor + 4 workers; supervisor holds routing logic, workers execute tools and return state via
add_conditional_edgesTradeoff
Supervisor on critical path of every step; per-step concurrency lost in exchange for auditability
Reversal
Peer-to-peer mesh is ~2 engineer-weeks if real-time streaming becomes a requirement
ADR-004Accepted
HITL via LangGraph `interrupt_before` + Slack actionable buttons
Context
Pause + persist + resume + audit + timeout-then-reject — message-queue + DB-flag patterns add too much infra; native graph pause wins
Decision
interrupt_before=[...] at compile time + Slack Approve / Deny / Escalate buttons → graph.aupdate_state resumesTradeoff
Slack on critical path; 24h hard timeout (ties to Redis TTL); approve/deny/escalate only — no partial state edit
Reversal
SQS-backed approver web UI is ~2 engineer-weeks; LangGraph contract unchanged
ADR-005Deprecated
Single global ToolRegistry without RBAC scoping (v0)
Context
Day-2 MVP: all agents shared one registry; nothing prevented Quality agent from calling Transform tools
Decision
Reverted in M03 — added
ScopedToolView with per-agent-role scopes (default-deny) at the registry layerWhy reversed
Quality agent silently applied a schema migration during a validation run on 2026-04-12 — broke M04 dashboard for 23 min
Replaced by
RBAC-aware registry; ~4.5 engineer-day reversal cost
EXPERT-only · cost model
Read the FinOps story, not just the latency one.
Module 04 ships a runnable cost-model CSV inside the starter-kit zip at docs/cost-model/. 5,000 agent-runs/mo load, real Anthropic + AWS RDS + ElastiCache + LangSmith list prices, with model-cascade and reserved-instance levers wired up. The version you’ll defend to a CFO. Preview the CSV →
ComponentBaseline / moOptimized / moDelta
Anthropic Claude Sonnet (planner + supervisor)
100% baseline → 30% optimized · 24M in / 6M out tok/mo
$144
$43
−$101
Anthropic Claude Haiku (worker agents)
70% of mix in optimized · ~17M in / 4M out tok/mo
$0
$26
—
AWS RDS Postgres (db.t4g.medium)
100GB gp3 · business data store (per ADR-002)
$50
$35
−$15
AWS ElastiCache Redis (cache.t4g.small)
checkpoint + recovery state (per ADR-002)
$35
$26
−$9
LangSmith observability (Plus tier)
30k traces/mo budget · agent + tool spans
$39
$39
—
GitHub Actions + container registry
~150 PR runs/mo × 6 min × Linux runners + GHCR
$12
$12
—
Total · 5k runs/mo
~$0.056 per run at baseline · ~$0.036 optimized
$280
$181
−$99 (−35%)
Optimization levers
Model cascade (Haiku for workers · Sonnet for supervisor)
Route 70% of worker LLM calls to Haiku ($0.80/M in). Supervisor + escalation paths stay on Sonnet for routing quality. ADR-001 + ADR-003.
−$75 / mo · −27%Idempotent tool-call cache
SHA-256 cache on (tool_name, args) for read-only tools. Redis-backed, 1h TTL on quality-sensitive, 24h on static reads. ~18% hit rate on regression suites.
−$14 / mo · grows with workload stabilityRDS + ElastiCache 1-yr reserved
Commit to 12-month reserved capacity once load is stable for 30 days. ~30% off RDS, ~26% off ElastiCache. Break-even at month 4.
−$24 / mo · −28% on store costEXPERT benefit · cohort beta
Async architecture review with a staff-level reviewer (cohort beta).
Submit your repo, your ADR draft, or your supervisor-routing prompt. A staff or principal-level reviewer who has shipped this exact stack responds within 7 days with line-by-line comments. Cohort capped at 12 members.
Bring a diff, an ADR draft, or an HITL flow.
The cohort beta runs as async architecture review — pick a reviewer by topic, send the artifact, get inline comments + a Loom walkthrough back. No back-and-forth scheduling. No 30-minute slot pressure.
MR
Mira R.
Ex-staff · agent platform · top-3 cloud
Multi-agent topology, supervisor design, RBAC at the tool layer, agent eval
“Send the diff. I'll go line-by-line through your supervisor prompt and the ScopedToolView wiring and pick out the agent-overreach paths.”
DK
Daniel K.
Principal · LLM platform · enterprise SaaS
HITL design, approval-flow auditability, time-travel replay, compliance + audit log
“Send your worst stuck-approval. We'll walk it backwards from the Slack event log to the LangGraph checkpoint state and figure out where the protocol broke.”
AS
Anya S.
Eng manager · AI platform · public Series-D
Org design for agent teams, hiring rubrics, staff-engineer interview prep, ADR review
“If you're prepping for staff promo, send your ADR draft. We'll work backwards from the rubric.”
What your tier unlocks
PRO unlocks Modules 01-03. EXPERT unlocks the full platform.
PRO is the entry point — Modules 01-03 (a working multi-agent pipeline) plus the rest of the PRO catalog. EXPERT unlocks Modules 04-06 of this build, the 5 ADRs, the cost-model CSV, and the cohort-beta async review.
What you getFREEPROEXPERT
Modules 01-03 of P13
Foundation + tools + multi-agent pipeline (~10h)
—
Included
Included
Modules 04-06 of P13
Production + hardening + platform design (~7.5h)
—
—
Included
5 committed ADRs + cost-model CSV
Starter kit docs/adr/ + docs/cost-model/
—
—
Included
PRO project catalog
Production-grade builds
2
All current
All current + this one
Curriculum
All 7 tracks
Phase 1 only
All
All + bonus modules
Code review
Senior+ reviewers
—
4 / month
Unlimited
Cohort-beta architecture review
Async · 7-day turnaround · 12-member cap
—
—
Included
Certificate
Verifiable on LinkedIn
—
Yes
Yes + LinkedIn rec
$79/mo
billed monthly · open enrollment · cancel anytime
or annual
$699/yr save 26%
Who this is for
Pick this if you own the supervisor, not just a feature.
ST
Staff / principal engineers · agent platform
You own the supervisor prompt, the RBAC boundary, and the answer to 'which agent decided this?' that your VP asks at the next incident review.
EM
Engineering managers · AI
You need a reference architecture for the agent platform your CTO will ask about before the AI team gets headcount or a budget for production deployment.
PA
Platform / infra leads
You absorb LangGraph without absorbing 4 new vendors. Postgres, Redis, Prometheus, Slack — tools you already operate. This is the playbook.
FR
Founding engineers · AI startups
Your investors will ask 'how do you know agents are safe to ship?' before they ask about scale. The 5 ADRs + HITL gate + RBAC registry is the answer.
Related curriculum
Going deeper? Four tracks back this project.
The Agentic Workflows curriculum is the foundation. These four tracks let you go deeper on agent eval, retrieval, production ops, and the platform-design discipline you'll need at staff level.
FAQ · EXPERT tier
Quick answers.
Modules 01-03 (the working multi-agent pipeline — foundation, 4 tools, supervisor + workers + Redis checkpointing) are included with PRO at $29/mo. The rest of the platform — Modules 04-06 (production observability, HITL + hardening, multi-tenant platform design), the 5 committed ADRs, the runnable cost-model CSV, and the cohort-beta async architecture review — unlocks with EXPERT at $79/mo. PRO gets you a working pipeline; EXPERT gets you the platform you'd defend in an architecture review.
Yes — most of the value is in the design decisions, not the framework. ADR-001 lays out exactly when LangGraph wins vs CrewAI / AutoGen / custom; the supervisor-worker topology in ADR-003 is framework-agnostic; the RBAC ToolRegistry in ADR-005 is a Python pattern. The orchestrator-specific code is contained to one directory (~200 lines) and is documented as a swap target.
Not for v1. The cohort beta runs as async review: you submit a diff / ADR / HITL flow / supervisor prompt, a staff-level reviewer responds within 7 days with inline comments + a Loom walkthrough. Cohort is capped at 12 members so reviewers can keep the SLA. We'll evaluate live 1:1 sessions once the cohort signal is solid.
17-18 hours of focused work across 6 modules. Most learners spread it across 4-6 weeks alongside a day job. Modules 01-03 alone are ~10 hours and get you a working multi-agent pipeline you can deploy locally.
It's a strong forcing function. Staff agent-platform interviews lean heavily on system design (multi-agent topology, HITL, audit, cost) and on having opinions backed by real tradeoffs. The 5 ADRs you commit (one Deprecated, with the RBAC reversal incident) are exactly the artifacts a panel asks about. Pair with the cohort-beta review on your final repo and you have the portfolio piece.
Yes — receipts and a learning-budget letter are downloadable on subscription. Many EXPERT learners are reimbursed under engineering training, AI upskilling, or platform-tooling budgets.
Model fine-tuning. Pre-training. RAG retrieval pipelines (we use them; we don't build them — see /projects/enterprise-rag for that). Agent training via reinforcement learning. This is an agent execution platform — you ship the system that runs agents in production, not the system that creates them.
Ready to ship the system that runs agents in production?
Start with PRO ($29/mo) for Modules 01-03 — the working multi-agent pipeline. Or unlock the full 6-module platform plus 5 ADRs, the cost-model CSV, and cohort-beta architecture review with EXPERT ($79/mo).
P13 · Agentic data pipeline · EXPERT · PRO unlocks M01-M03Unlock EXPERT →