Video summary
AWS re:Invent 2025 - Building software like never before with Agentic AI (DVT220)
Main summary
Key takeaways
Summary of technological concepts, product features, and analysis
Session purpose and themes (Agentic AI for software development)
- The talk frames Agentic AI as a shift from:
- code prediction / autocomplete → chat-based assistance → agents that understand context, converse, and autonomously take actions
- Emphasizes that AI dev tooling is changing rapidly.
- Notes that many teams already use multiple tools, but production readiness and control remain major concerns.
AWS Kiro: “Agentic development environment” for production-grade structured AI coding
AWS introduces Kiro (launched Nov 17; GA “couple of weeks ago,” preview in July) as an environment designed to apply mature engineering practices to AI coding—aiming at production-level applications using structured/spec-driven workflows.
Key motivations / pain points (from customer feedback)
- Autonomy limits in “vibe coding”
- Great for small tasks/prototyping, but production work still requires substantial follow-up (bug fixes, code edits, rewrites).
- Limited control & visibility
- Some tools lack monitoring/end-to-end transparency and the ability to intervene during agent actions.
- Quality & verification challenges
- Agents can produce code/tests/docs, but teams need intuitive ways to confirm outputs match developer intent and project standards.
Core product features described
Two interfaces
- Kiro IDE (fully agentic development experience)
- Kiro CLI (agentic terminal workflow)
Spec-driven development (vs. generating large code dumps early)
Kiro emphasizes a guided workflow that generates:
- Requirements
- user stories + acceptance criteria
- Design artifacts
- components/interfaces, API routes, error handling, testing strategy, etc.
- Task list / MVP approach
- task backlog tied back to the spec
Additional capabilities:
- Requirements traceability
- mapping requirements → tasks
- Human-in-the-loop edits and iterative refinement before coding
Demo example: Flask event management app
- Prompt: add a “category” feature and enable event filtering.
- Kiro generates:
- requirements (user stories + acceptance criteria)
- design markdown (components/interfaces/routes/error handling/testing strategy)
- a task list, then code
- Showcased:
- approvals for agent commands
- automated testing
- database/table changes
- iterative task completion
Agent hooks (automating repetitive dev workflows)
Kiro includes agent hooks to trigger actions on events such as:
- file save
- Python file updates
Examples mentioned:
- regenerate test cases
- update documentation
- update changelog (timestamp + summary)
- propagate changes across multiple languages/labels (example described)
Advanced context management with “steering files”
To reduce mismatches with org standards/best practices, Kiro supports steering files, such as:
- project coding standards
- testing guidance (libraries, strategies)
- security guidelines
- deployment process rules
Steering files can integrate with spec-driven development—for example, “persona” files used when generating requirements.
Brownfield steering demo
Kiro can generate/support steering docs like:
- product markdown (core features, UX, business logic)
- tech markdown (technology + DB schema)
- structure doc (file/module structure)
It can also add:
- coding standards (including recommendations produced by Kiro)
The approach emphasizes opinionated templates to accelerate brownfield onboarding.
Announcements mentioned
- Kiro powers
- partner integrations (example: Figma)
- Frontier agent
- a virtual extra teammate for Kiro
Rackspace customer story: enterprise rollout and outcomes using Kiro
A customer (Rackspace) describes evaluating and adopting Kiro as part of a standardized “one Rackspace” development community.
Evaluation approach
- They saw other AI coding tools as strong in parts but lacking cohesive coverage across needs.
- Kiro was selected due to its strengths and evolving capabilities, becoming the preferred tool.
Reported results / metrics
- Early claim: 81% overall efficiency gains (updated to 85%).
- Scale:
- 700 developers (later stated 800 developers total)
- 51+ projects logged
- Estimated time savings: 8+ FTE years across 4–5 months
- Business framing:
- shifted from perpetual maintenance to continuous delivery & innovation
- improved security and compliance
Example use cases
- Legacy Python application modernization/upgrade
- Added missing unit tests and documentation
- Identified dead/redundant code
- Reduced code size (“shaved thousands of lines”)
- Improved maintainability/security/reliability
- Reported: first app moved to production after using Kiro
- MCP server + Splunk integration for post-deploy error remediation
- Integrated Kiro’s MCP server with Splunk to:
- deduplicate logs
- identify errors after go-live
- Because context stayed tied to the codebase in Kiro, the team could quickly:
- write missing unit tests
- fix defects
- redeploy
- Reported: 3 defects fixed within 15 minutes with no deployment complaints (for a critical internal system)
- Integrated Kiro’s MCP server with Splunk to:
Rollout methodology
- “Start small and controlled”
- initial spark/ignition cohort (8 teams; multiple languages including Python, Java, .NET)
- playbook included modernization steps like adding tests and creating spec files
- Emphasized communication cadence: “talk early, talk often”
- Expanded users: 100 → 175 → eventually to 800
- Created a Community Center of Excellence
- recurring sessions every other Friday with speakers sharing innovations and practices
- Goal:
- reduce IT bottlenecks and empower business units and cross-functional teams with a shared use-case library
Building custom agents inside applications (AWS enabling production agents at scale)
The second half shifts to agent design embedded in business applications.
Why build agents into apps (analysis)
- Agents make natural language a UI, enabling users to request actions directly.
- Improve customer experience via personalization and memory/reasoning.
- Provide context-aware intelligence, reducing the need for massive bespoke code paths and long-term maintenance of customer-specific workflows.
- Competitive framing: assume competitors are adopting agent-based automation.
Gartner predictions cited
- Generative AI in a third of enterprise applications by 2028 (up from ~1% in 2024)
- At least 15% of work decisions made autonomously via generative AI by 2028
AWS “Agent Core” approach: optionality + production primitives
The talk emphasizes AWS’s goal: let developers choose tools/models while still achieving production-grade deployment.
Optionality (developer tooling + model choice)
- No one wants to lock into a single framework/model for 10 years.
- AWS positions itself to support multiple frameworks and models.
Components referenced
- Strands Agent SDK (open-source)
- Build agents from a prompt + tools list
- test locally
- deploy to cloud
- Amazon Bedrock
- “single API” access to 100+ models
- Amazon Bedrock Agent Core (modular services for production)
- Runtime
- serverless deployment
- session isolation
- long-running low-latency agent work
- Memory
- short-term and long-term memory capabilities
- Identity/Access
- controls what AWS resources/tools the agent can access
- Gateway / Browser tool
- connect agents to MCP/openAPI/Smithy protocols
- interact with web pages via computer-use browser runtime
- New primitives mentioned:
- Evaluations primitive
- benchmark agent performance against metrics/business goals
- Policy primitive
- described as another control primitive for production behavior
- Evaluations primitive
- Runtime
Demo: customer support agent workflow
Scenario: customer requests a game exchange immediately.
- “Old experience”: delayed email
- “New experience”: agent processes in real time
High-level steps shown:
- In Kiro:
- create steering documents
- use spec mode to define agent requirements
- Integrate with Strands + Agent Core MCP tools
- Generate:
- requirements → design → task list → source code
- Deploy via an Agent Core configure command to Agent Core runtime
- Run a test prompt (“hello”) to validate the agent is live
Observability emphasis
- Uses OpenTelemetry for observability.
- Demonstrated a “decision mapping tree” view to explain why the agent made decisions and support iteration when outputs don’t match team expectations.
Main speakers / sources
- Svetlana Kolomeysky (AWS) — moderator; leads outbound go-to-market for Agentic AI development experience services
- Al Di Stefano (AWS) — senior go-to-market specialist; Agentic AI developer experience team
- Brian Lickley — Chief Technology Officer, Rackspace IT
- AWS / Gartner — referenced for market predictions (Gartner cited in the second segment)