AgentCore course → Lesson 6

Lesson 6· ~12 min· Reading + retrieval practice

What's new, and the frontier

AgentCore's newest surface area is where customers will test whether you are current. It is also where the temptation to over-promise is strongest, because two of the three most interesting things here are still preview.

Start from what you have

You've sat on the other side of this: a vendor demos something exciting, you ask "is that GA?", and the answer arrives three sentences later than it should. Your credibility with a technical founder rests more on how fast you volunteer that answer than on the feature itself.

harness, properly

Lesson 1 introduced harness as the managed-loop alternative to Runtime. It reached GA on 17 June 2026 and deserves a closer look, because it is the most significant change to AgentCore's shape since launch.

AWS's framing is worth learning verbatim: an agent is more than a model. If the model is the brain, the harness is the body. It runs the orchestration loop, executes tools, manages the context window, persists state across turns, recovers from failures, and isolates each session. AWS goes further and claims the harness shapes how well an agent performs as much as the model does, and that building a durable one is where most teams spend their time today.All harness detail in this section from AgentCore harness is now generally available, 17 June 2026, and AgentCore adds new features to help developers build agents faster, April 2026. Retrieved 2026-08-05.

What you actually get from a single configuration: a production-grade agent running in minutes in its own isolated environment with a filesystem and shell, memory across sessions, skills including an AWS-curated catalog, and web browsing. Four properties are worth carrying into a customer conversation:

Model-decoupled, mid-session

Switch model providers mid-session without losing context: plan with one model, write code with another.

The most concrete anti-lock-in fact in AgentCore.

Override per invocation

Any create-time configuration can be overridden on a single invocation, so a team experiments without redeploying.

Changes the iteration loop from minutes to seconds.

Not a starter tool

AWS's explicit claim: the configuration you start with is what you operate at scale, or as AWS puts it, "the agent declared on day one is the agent that runs at the thousandth."

Pre-empts "we'll outgrow this."

Exports to code

One CLI command exports the harness orchestration to Strands-based code, on the same compute and the same primitives. Claude Agent SDK is a stated future target.

This is the door back out.

Two entry paths onto one platform. Writing configuration leads to harness, where AWS assembles and runs the loop. Writing orchestration code in LangGraph, CrewAI, Strands or ADK leads to Runtime, which hosts the loop you wrote. Both reach identical primitives: Gateway with Policy, Identity, Memory, and Observability. A single CLI command exports a harness configuration into Strands code on the Runtime path.
Figure 1. Two doors, one platform, and a way back. Configuration leads to harness; code leads to Runtime. Both land on identical primitives, so the choice is about who owns the loop and nothing else. The export command turns what looks like a one-way door into a reversible decision, which is why "start with harness" is safe advice.

Open full-size diagram

The one idea

Because harness exports to code on the same primitives, choosing it is reversible. That single fact is what makes "start with harness and see" a responsible recommendation to a startup rather than a bet on a managed abstraction.

Two supporting facts worth having ready. harness, the AgentCore CLI, and skills carry no additional charge, so you pay only for the underlying capabilities consumed. And the CLI deploys with infrastructure-as-code governance, with AWS CDK supported today and Terraform announced as coming. There is also a preview filesystem persistence capability that externalises session state so an agent can suspend mid-task and resume exactly where it left off.

AWS Agent Registry, and a live deadline

Registry is a single place to discover, share, and reuse agents, MCP servers, tools, and agent skills across an organisation, with governed publish-review-approve workflows and combined semantic and keyword search. Record types at GA are AGENT, MCP, SKILL, and CUSTOM.Registry scope and record types, AWS Agent Registry; migration detail from the migration guide. Retrieved 2026-08-05.

The problem it solves is one your enterprise-adjacent customers will recognise immediately: once several teams build agents independently, nobody knows what exists, two teams wrap the same API differently, and no one can answer "which agents can reach production data?" Registry is the answer to agent sprawl.

Time-critical: check this before you quote anything

Registry reached GA on 6 August 2026, one day after these lessons were written, and moved from the bedrock-agentcore namespace into its own agent-registry namespace with breaking API schema changes.

If you have a customer on the Registry preview, this is the single most valuable thing you can tell them this month. Full detail on the service map card.

Payments and x402: genuinely new territory

Payments Preview lets agents autonomously access and pay for APIs, MCP servers, web content, and other agents. Built with Coinbase and Stripe, it handles the full lifecycle from wallet authentication through transaction execution to spending governance and observability. Agents that transact: AgentCore now includes Payments (preview), April 2026. Retrieved 2026-08-05.

The mechanism is elegant enough to be worth explaining precisely, because it is the part that makes engineers lean in. An agent encounters a paid resource and receives an HTTP 402. AgentCore then handles the x402 protocol negotiation, wallet authentication, stablecoin payment, and proof delivery back to the endpoint, without interrupting the agent's reasoning loop. Spending limits are enforced deterministically at the infrastructure layer, not in the prompt, and every transaction is observable through the same traces.

Two concrete details worth carrying: wallets are Coinbase CDP or Stripe Privy, and the Coinbase x402 Bazaar MCP server is available through Gateway, exposing over 10,000 x402 endpoints an agent can search, discover, and pay for autonomously.

Note the pattern repeating from Lesson 3: spending limits enforced deterministically at the infrastructure layer is the same architectural argument as Policy at the Gateway perimeter. AgentCore's consistent position is that constraints on autonomous systems belong outside the model's reasoning, not inside its instructions. If you internalise one design philosophy from this course, make it that one.

Canada consequence

Payments is not in ca-central-1. N. Virginia, Oregon, Frankfurt, and Sydney only. For a Canadian agent-native startup interested in agent-to-agent commerce, this means an out-of-region deployment for the payments layer. Given it is preview, that is usually acceptable for experimentation but not for a production roadmap.

How to talk about preview without over-promising

This is a skill, not a disclaimer. Four rules that keep you credible:

  1. Volunteer the status before you are asked. Say "this is preview" in the same breath as the feature name. Waiting to be asked reads as concealment even when it isn't.
  2. Separate "can pilot" from "can plan." Preview capabilities are usually fine to experiment with and wrong to put on a funded roadmap. Optimization being free in preview makes it genuinely cheap to try, which is a better pitch than implying it is finished.
  3. Name what could change. AWS's own docs warn that preview features and APIs may change before GA. Registry is the live proof: it changed namespace, endpoints, IAM prefix, and API schema at GA. Customers who built on the preview have real migration work.
  4. Never guess a GA date. Optimization and Payments have no announced GA timing. "I don't know, and I'll find out" costs you nothing; a wrong date costs you the relationship.
Say this to a customer

"Two of the things I just showed you're preview: Optimization and Payments. That means pilot them, don't plan around them. Registry just went GA and if you're on the preview you've a migration with a September deadline. Everything else I've described is generally available."

One aside relevant to you specifically: AgentCore ships skills for coding assistants that provide accurate, up-to-date AgentCore guidance, distributed through Kiro Power with Claude Code, Codex, and Cursor support announced. If your SA team uses any of these, that is a faster route to current AgentCore knowledge than documentation.

Read this next: 5 minutes, and the best-argued of the set

AgentCore harness is now generally available

Unusually well written for a What's New post. Contains the brain-and-body metaphor, the mid-session model switching claim, and the export-to-Strands escape hatch. Those are the three facts that do the most work in a lock-in conversation. Retrieved 2026-08-05.

Retrieval practice

Four questions on maturity and the newest surface area.

  1. Which pair of capabilities is still in preview?

    • The harness and Evaluations, both announced in 2026
    • Registry and Memory, both recently restructured
    • Optimization and Payments, both without a GA date

    Answer: C (Optimization and Payments).

  2. A founder worries that starting with harness traps them in a managed abstraction. What is the accurate reassurance?

    • The configuration format is portable to any framework
    • One CLI command exports it to Strands code
    • AWS contractually guarantees the config never changes

    Answer: B. It exports to Strands code on the same primitives.

  3. What triggers the x402 flow in AgentCore Payments?

    • The agent receives an HTTP 402 from a resource
    • The agent exceeds its configured spending limit
    • The developer configures a wallet on the gateway

    Answer: A (an HTTP 402 response).

  4. What design philosophy do Policy and Payments spending limits share?

    • Constraints are expressed as careful prompt wording
    • Constraints depend on which model the agent uses
    • Constraints are enforced outside the model's reasoning

    Answer: C. Enforced at the infrastructure layer, outside the reasoning.

Where to go next

You are now current on AgentCore's full surface area and can give a straight answer on maturity for every capability. Lesson 7 is the drill: three Canadian startup scenarios where you pick primitives, justify them, cost them, and handle a planted objection.

This lesson produces the objection-handling card, covering five objections you will actually hear, with honest answers and the real remaining gap in each.

Questions to take forward