Healthcare
Claims Processing Agent
A document-heavy claims workflow where policy rules, human review, and a complete audit trail are the product, and the model is the smallest part.
documentspolicyhuman review
COURSE 001 · FDE ENGINEERING
Engineering Long-Running AI Agents
Learn how to design agents whose work survives failures, restarts, long waits and changing context.
Course content and API examples last verified: 2026-08 · exact dependency versions live in the reference repo lockfile
What you'll learn
- explicit durable state, with evidence instead of claims
- checkpoints, replay semantics, and recovery
- replay-safe steps and what may execute twice
- idempotent side effects and reconciliation
- worker ownership: claims, leases, heartbeats, fencing, orphan recovery
- failure-specific retry policies, timeouts, and honest degradation
- durable human approval that never holds a worker
- artifact-based context management and handoffs
- traces, events, and an operator dashboard
- security boundaries untrusted content cannot cross
- trajectory evaluations and validated judges
- a Deep Agents implementation, interrogated against your invariants
- a deployable runtime with startup reconciliation and a runbook
Industry cases
You build one canonical system, the Vendor Review Agent, all the way through. The concepts are grounded in five industry cases along the way: incident investigation, claims and document workflows, customer operations, enterprise data, and coding agents. One app to build, five systems to understand.
SaaS
Customer Account Agent
An operations agent that acts on customer accounts through real APIs, where every action is a side effect someone has to approve, dedupe, or reverse.
tool permissionsside effectsidempotency
Financial Services
Enterprise Data Agent
A question-answering agent over enterprise data where the hard problems are permissions, business definitions, and proving the answers are right.
context engineeringdata permissionsSQL
Developer Tools
Repository Migration Agent
A coding agent that migrates a large codebase over many sessions, with plans, parallel workers, and verification that does not take "I finished" on faith.
coding agentsplanningrepository context
Enterprise Operations
Incident Response Agent
A production investigation agent that fans out across logs, metrics, deploys and tickets, survives partial failure, and synthesizes a defensible incident narrative.
parallel executionpartial failurelong-running agents
Course modules
Each module answers one production question. Watch or read to follow the course completely; the labs and the repo are optional depth.
What Is a Long-Running Agent?
When is an agent worth this much engineering?
Video coming soon Read lesson
Build the Naive Agent
Forty lines that work. Every line is a promise you cannot keep.
State: What Must Survive?
The agent says it finished. Prove it.
Durable Execution, Checkpoints, and Replay
Your worker disappeared. Did the job?
Idempotency, Side Effects, and Reconciliation
The customer got two refunds. The agent swears it sent one.
Work Ownership: Claims, Leases, Heartbeats, and Orphans
Two workers picked up the same investigation.
Failure Handling, Retries, Timeouts, and Degradation
The log store is down. Should the investigation die?
Human-in-the-Loop and Durable Waiting
The approver is at lunch. For three hours.
Context Engineering Across Long Runs
The agent read forty pages and got dumber.
Observability and Run UX
The run looks dead. Is it?
Security: The Hostile Vendor
The vendor's website is talking to your agent.
Evaluating Long-Running Agents
Version B feels better. Feels.
Video coming soon Read lesson Explore implementation
Optional lab 11: The Judge That Agreed With Everyone
The Deep Agents Bridge
The framework says it handles all this.
Industry Architectures and Case Studies
Every credible system has the same skeleton.
Video coming soon Read lesson
Deploying and Operating the Agent
Push to main, mid-run.
Production Failure Gauntlet
Everything fails at once. That is the exam.
Technologies used
Python
The whole reference implementation; modules 00-02 are plain Python so the loop is visible.
LangGraph
Graph execution, persistence, checkpoints and interrupts, adopted in Module 03 after the failures make them necessary.
PostgreSQL
Durable state, leases, event log, side-effect records. The load-bearing component.
FastAPI + HTMX
The run API, review endpoints and operator dashboard, with no frontend framework.
Render + Neon
The free-tier public deployment whose restarts double as failure injection.
Optional labs
14 Failure Labs break the running system on purpose: kill the worker, duplicate the publish, poison a page, race two workers. Entirely optional; the lessons describe each failure well enough to learn from without running anything.
Reference implementation
The Vendor Review Agent lives in an open repo with the fixture server the labs run against. Clone it if you want your hands on the system; skip it if you are here to watch and read.
Additional resources
Every module links one to three primary sources with a sentence on why each is worth your time. The full annotated list lives on the course sources page, and the site-wide library is at /resources.
Local progress only: your Watched/Read marks and exit-criteria checklists live in this browser, never on a server.