Day 3 of re:Invent brought one of the most thoughtfully constructed keynotes I’ve seen from AWS. Instead of launching straight into features, Dr. Swami Sivasubramanian started with the actual challenges we all deal with when building AI systems: hallucinations, fragmented tooling, complex infra, and the difficulty of putting agents into real workflows.
From there, he walked through how AWS is solving these problems — one layer at a time — and immediately showed each solution in action with engineers and customers. The result was a keynote that felt honest, practical, and very builder-focused.
Agents, Not Chatbots
Swami made a clear distinction early on:
Chatbots answer. Agents act.
Agents:
- Investigate
- Diagnose
- Plan
- Call tools
- Execute
- Check results
- Learn over time
He broke an agent down into three components — the model, the code, and the tools — and somehow made the architecture of agentic systems feel very approachable.
Strands Agent SDK — Now for Everyone
One of the biggest unlocks was the evolution of the Strands Agent SDK.
Key updates:
- TypeScript support, massively opening the door for web developers
- Edge device support for robotics, automotive, gaming, and more
- A more streamlined, model-driven approach to building agents with minimal code
This is the moment where agentic AI moves beyond ML circles and into mainstream development.
Amazon Bedrock AgentCore — The Runtime Agents Needed
AgentCore was easily one of the most impactful announcements.
It tackles everything that makes production agents hard:
- Memory systems (short-term, long-term, episodic)
- Identity and permissions
- Tool orchestration
- Observability
- Multi-step execution
- Deployment and scaling
And episodic memory really stood out — agents can now remember past experiences, learn from them, and adapt over time.
This is the beginning of agents that actually get better the more you use them.
Model Customization — Accessible to Every Developer
Swami spent a big chunk of time on customization, but for the first time, the entire stack felt simple and usable.
Techniques he covered:
- Supervised Fine-Tuning (SFT)
- Model Distillation
- Reinforcement Learning (RL) with human or AI feedback
- Reinforcement Fine-Tuning (RFT) on Bedrock
RFT was the standout:
It automates reward modeling, evaluation, training, and safety checks — something that traditionally needed research-level ML teams.
This is AWS taking advanced model engineering and making it accessible.
SageMaker AI — Full Control When You Need It
While Bedrock focuses on simplicity, SageMaker AI remains the place for deep control.
Swami highlighted:
- End-to-end pipelines for training custom models
- Full ownership of data, methods, and infrastructure
- A new serverless customization experience guided by agentic workflows
Bedrock and SageMaker now feel intentionally complementary:
Speed vs. control.
Nova Forge — Build Your Own Frontier Model
This one surprised me.
Nova Forge gives enterprises the ability to build their own frontier-scale models by:
- Accessing intermediate checkpoints
- Mixing proprietary data into Nova’s training cycle
- Retaining Nova’s safety and foundational knowledge
- Significantly reducing the cost of training from scratch
This is frontier model development without needing frontier-scale resources.
SageMaker HyperPod & Checkpointless Training
HyperPod continues to mature into a powerhouse for training large models.
What really changed the game was checkpointless training:
- The model state is continuously preserved across the cluster
- Node failures automatically recover in minutes
- No manual checkpointing
- No lost training runs
- Large-scale training becomes more efficient and cost-effective
This solves one of the most painful problems in LLM training.
Real Customer Stories — Grounding Everything
Hearing from teams like Vercel, Blue Origin, and others helped connect the announcements to real-world use cases.
Examples included:
- Developer workflows powered by agentic tools
- Production use of Bedrock and Nova
- Reinforcement learning applied at scale
- Real latency and cost improvements
These weren’t theoretical demos — they were practical, in-production stories.
A Glimpse Into the Future
Swami closed the keynote by painting a future where agents:
- Handle complex, multi-step workflows
- Reason about tasks and coordinate with each other
- Become more reliable through episodic memory
- Operate across cloud, edge, and devices
- Are trained with proprietary industry knowledge
- Move toward autonomous systems
It didn’t feel like hype. It felt like a roadmap.
Nova Act — The Final Launch
Right before wrapping up, Swami introduced Nova Act — a service for enabling highly reliable agent actions inside web browsers.
This unlocks:
- Workflow automation
- Browser-based RPA
- UI-heavy enterprise workflows
- Agents that actually take actions, not just provide instructions
It ties the entire agentic stack together.
Celebrating Builders — Hackathons, Programs, and Winners
After the announcements, Swami shifted to something that genuinely made the keynote more personal — recognizing the developers who make all of this possible.
He highlighted:
- Ongoing developer-focused hackathons
- Community-driven agent-building initiatives
- Global innovation programs leading up to re:Invent
- Grassroots developer momentum around agentic AI
He also congratulated the winners of:
- Road to re:Invent
- AI League
It was a nice moment — a reminder that behind all this powerful tech are actual builders pushing the boundaries.
Final Thoughts
This keynote wasn’t about flashy announcements. It was a blueprint for the next era of AI — an era where everyday developers can build powerful, reliable, and intelligent agents without needing a PhD.
For me, the biggest shift is clear:
AI is moving from answering questions to solving real problems.From chatbots to agents.From models to systems.
And AWS is quietly building every layer to make that future real for all of us.
If you want a more unfiltered, real-time version of my thoughts during the keynote, I also live-tweeted the entire session on X. It includes reactions, screenshots, and moment-by-moment highlights that didn’t make it into this blog. You can check out the full thread here: https://x.com/adi_12_modi/status/1996257235731955907?s=20
Top comments (0)