PRIMARY SOURCES

Read the originals

Wherever the course makes a framework-specific or fast-changing claim, it cites a primary source. Modules link one to three of these in place; this page is the full annotated list. No link dumps: each entry says why it is worth your time.

Framework and hosting pages change quickly. These links are reviewed before each course revision, and the repo pins the exact dependency versions the course was tested against.

LangGraph and Deep Agents

LangGraph persistence
The authoritative answer to what one checkpoint contains and which tables the Postgres saver writes. Read next to your own checkpointer from Module 03.
LangGraph interrupts
Exact semantics of interrupt() and Command(resume=...), including what replays on resume. The replay behavior is the part that surprises people in Module 07.
LangGraph graph API
Where the docs themselves tell you to keep node logic idempotent, which is the bridge into Module 04.
Deep Agents overview
The harness Module 12 interrogates. Read the mechanisms, then map each to the module where you built it by hand.
Deep Agents context engineering
Module 08's three techniques as framework features.
Deep Agents subagents
Clean-context subagent spawning, the deep-dive pattern formalized.
Deep Agents going to production
Compare its defaults against your own security_policy.md from Module 10.

Long-running harness engineering

Anthropic: effective harnesses for long-running agents
The industry statement of the harness/runtime split this course is built around.
Anthropic: effective context engineering for AI agents
The published basis for Module 08's claim that compaction alone is not sufficient.
Anthropic: harness design for long-running apps
Harness design decisions for multi-session application development.
Anthropic: scaling managed agents
What running other people's agents as a service forces you to make durable.
Anthropic: multi-agent research system
A production long-horizon system described with unusual candor; Module 13's first extraction case.

Evals

Hamel Husain: LLM-as-a-judge
The methodology Module 11 follows, including judge validation against human labels. If you read one link on this page, read this.
Hamel Husain and Shreya Shankar: evals FAQ
Short answers to the exact objections your team will raise when you propose error analysis before metrics.

Durable workflow comparison

Temporal AI integrations
The landscape when a workflow engine owns durability instead of your tables.
Temporal durable OpenAI Agents example
Claim, lease, heartbeat and reclaim as engine primitives; compare with your Module 05 scheduler.
Temporal human-in-the-loop example
The durable-wait pattern of Module 07 in a workflow engine, showing it is an industry invariant and not a framework quirk.

Current industry examples

OpenAI: harness engineering
The coding-agent harness architecture from the team shipping it.
OpenAI: unrolling the Codex agent loop
The agent loop of Module 01, at production scale.
OpenAI: run long-horizon tasks with Codex
Long-horizon task guidance for a shipping agent product.
Replit: evaluating and improving Agent at scale
Evals as the development loop, argued by someone with millions of runs.
Replit: decision-time guidance
Steering an agent mid-trajectory rather than only at the prompt.
Replit: inside the snapshot engine
Durable execution machinery from a production agent platform.

Hosting limits

Render free instance limitations
The constraints Module 14 designs around. Verify before relying on the module's numbers.
Neon free-plan limits
Connection and storage caps that shape the lease scheduler and pooling choices.