Key talking points:
- Short-term memory is per-session — raw conversation events, up to 365 days retention
- Long-term memory is cross-session — async extraction via strategies: UserPreference, Semantic, SessionSummary, Episodic
- "No DynamoDB tables" is the pitch — you don't manage storage, indexes, or TTLs
- Multi-agent memory sharing is important for supply chain, support escalation, or any workflow where agents hand off
- Episodic memory (added Dec 2025) captures structured episodes: scenario, intent, actions, outcomes, reflections
- Summarisation uses advanced prompt templates and runs async — may take a few minutes after session ends
Key talking points:
- This is NOT just "give the agent an IAM role" — it's a proper workload identity system
- Token vault encrypts at rest AND in transit, validates every access independently
- 3LO (authorization code) is the flow where a human consents — agent gets scoped tokens, not your password
- 2LO (client credentials) is machine-to-machine — agent authenticates as itself
- Identity propagation means downstream services see BOTH the agent identity AND the user identity
- Declarative annotations in the SDK mean you don't write OAuth boilerplate — just annotate which credentials you need
- Cross-service agents (agent calling agent) maintain identity chain — no "confused deputy"
Key talking points:
- MicroVM isolation (Firecracker) means one session cannot leak into another — full CPU/memory/filesystem separation
- Live View is a React component you can embed — uses AWS DCV (same tech as WorkSpaces) to stream the browser
- Playwright is the automation layer — same API devs already know from testing
- Session recording to S3 is great for compliance — replay exactly what the agent did
- Web Bot Auth reduces CAPTCHAs — not eliminates — so agents can navigate sites that would normally block bots
- Browser profiles let you persist cookies/settings across sessions (e.g. logged-in state)
- OS-level actions mean the agent can do things like file system operations on the VM, not just browser tabs
- Enterprise policies let you restrict which sites the browser can access
Key talking points:
- Not just Python anymore — JS and TS are supported too
- Pre-installed libraries mean agents don't waste time pip installing numpy/pandas/etc
- 8-hour execution ceiling matches Runtime — good for batch data processing
- Direct invocation (no agent framework) is useful for "just run this code" use cases
- Code execution results stream back — you get incremental output, not just final result
- Containerized, NOT microVM like Browser/Runtime — different isolation model
Key talking points:
- "Few lines of code" to go from local agent to cloud deployment — that's the pitch
- Firecracker microVMs are the same tech as Lambda — proven isolation model
- 8-hour sessions are a differentiator — most serverless has 15-min limits
- Persistent filesystem is unique — agent can install packages, write files, and they survive across invocations within a session
- A2A protocol support means agents can call other agents natively
- Scales to zero means you're not paying for idle — important for dev/test
- Shell execution lets agents run arbitrary commands in their sandbox — powerful but needs policy guardrails
- ADOT SDK integration for custom tracing/metrics out of the box
Key talking points:
- Think of it as a "service catalog for agents" — who built what, where is it, is it approved
- Currently in Preview (April 2026) — 5 regions: us-west-2, us-east-1, eu-west-1, ap-northeast-1, ap-southeast-2
- Auto-discovery is smart — point it at a live MCP server URL and it pulls metadata automatically
- Hybrid search means "find me something that can send Slack messages" works alongside exact keyword matches
- EventBridge integration lets you plug into existing approval pipelines (ServiceNow, Jira, etc.)
- MCP-native endpoint means agents can discover and invoke tools from the registry directly
- Custom JWT auth is useful for multi-tenant SaaS platforms
- Deprecated records can be hidden from search but preserved for audit
Key talking points:
- "Out of the box" is key — deploy to Runtime and you get session metrics, spans, and logs automatically
- ADOT (AWS Distro for OpenTelemetry) SDK lets you add custom instrumentation — e.g. track business-specific metrics
- Spans cover the full chain: agent reasoning > tool selection > Gateway call > tool execution > response
- Works for agents NOT on AgentCore Runtime too — just configure ADOT SDK and point at CloudWatch
- Transaction Search in CloudWatch lets you trace a single request across agent > gateway > tools > memory
- Multi-destination logging means you can ship to S3 for long-term retention or Firehose for real-time analytics
- Policy spans show you what was evaluated and whether it was allowed/denied — useful for debugging access issues
Key talking points:
- Online mode continuously samples live traffic — set a percentage or filter by criteria
- On-demand mode is for CI/CD or investigating specific interactions — you pick the exact spans/traces
- Custom evaluators are the power feature — write a Lambda function with your own scoring logic
- Ground truth evaluators include trajectory matching: did the agent call the right tools in the right order?
- Three trajectory modes: exact order, in-order (allows extras), any-order (just checks presence)
- GoalSuccessRate evaluator takes natural language assertions — "the agent should have booked a window seat"
- Session-level vs trace-level matters: trace = one turn, session = entire conversation
- Built-in evaluators have public ARNs — custom evaluators are private to your account
- Each evaluator has its own ARN — can scope IAM permissions per evaluator