Amazon Bedrock AgentCore
A 200–250 level course for a Solutions Architecture Manager leading the AWS Canada startups team. Built to be read in short sittings, to hold up in front of a technical founder, and to be handed to an SA team as-is.
The credibility bar
Enough to hold an advisory conversation without overclaiming, and to know which question to take away rather than answer.
Add the tool-call path
The three patterns are simple enough to redraw from memory. Practise them before you need them in front of a founder.
- The 30-minute route, plus Lesson 3
- Whiteboard patterns — redraw each one
Start Lesson 1 Seven lessons, about 80 minutes in total. The lessons build, so read them in order.
What this course is for
Every lesson exists to make one specific customer conversation go better, which is why it skips a good deal of AgentCore. Every factual claim carries a citation to a primary AWS source with a retrieval date, because this service surface changes monthly and a confidently wrong answer costs more than an honest "I'll find out."
Lessons
- 01 Translating what you already know Where AgentCore sits relative to LangGraph and CrewAI, and why that framing decides every later conversation. ~10 min
- 02 Runtime, isolation, and the compute decision Session isolation, the Runtime / harness / Lambda / ECS-EKS decision, and what actually runs in ca-central-1. ~12 min
Meeting-credible from here
- 03 Gateway, Identity, and Policy The tool-call path end to end, and why the gateway is where governance becomes enforceable. ~11 min
- 04 Memory and context Short-term versus long-term memory, the three billing meters, and the strategy choice that moves cost 3x. ~10 min
- 05 The quality loop Observability, Evaluations, and Optimization — how you answer "how do you know it's any good?" ~11 min
- 06 What's new, and the frontier The harness in depth, Agent Registry, and Payments over x402 — plus how to discuss preview honestly. ~12 min
- 07 Advisory whiteboard drill Three Canadian startup scenarios. Pick primitives, justify, cost it, handle the objection. ~15 min
Each lesson ends with retrieval practice: multiple choice with immediate feedback, and you can retry a question you get wrong. Answer from memory rather than scrolling back — getting one wrong is more useful than getting it right by looking.
Field cards
The compressed residue of the lessons, sized to print and carry. These are what you revisit; the lessons you will probably read once.
- Card Canada & data residency What runs in ca-central-1, and where prompt data actually travels. Almost nobody else in the meeting will have this one, and there is a finding in it that changes how you advise regulated customers. 2 sides · from L2
- Card Service map Every capability, its maturity, and its region footprint. 1 side · from L2
- Card Pricing cheat sheet Every meter and rate, plus worked cost sketches at startup scale. 2 sides · from L2
- Card Whiteboard patterns Three reference architectures, one per page, simple enough to redraw in ninety seconds. 3 sides · from L7
- Card Objection handling Five objections you will hear, with honest answers and the real remaining gap. 3 sides · from L6
- Card Terminology card AgentCore's confusable names, disambiguated. Read this first if a term trips you up. 2 sides · from L1
This course has a shelf life
Facts were verified on 5 August 2026. AgentCore shipped the harness to GA in June 2026, Evaluations in March, Optimization to preview in May, and Agent Registry to GA on 6 August — the day after this was written. Treat anything here older than about six weeks as suspect, and re-verify against the region matrix and the AWS What's New feed before a meeting that matters.
Asking me to re-verify the course and report what changed is a faster path to current than rereading it.
Questions to take forward
- Work a real customer workload through the whole course with their constraints.
- Role-play a sceptical CTO on any lesson in here.
- Go deeper on anything that felt thin, or come at it from a different angle if it did not land.
- Re-verify the course against AWS primary sources and tell me what changed.