The AWS Developers Podcast

John Sexton and Aaron Tummon from Genesys join the show to share hard-won lessons from building and operating Cloud Copilot — an agentic AI layer serving 2 million users across 21+ AWS regions. Genesys powers customer experience for brands like Virgin Atlantic, Vodafone, and HSBC, and their copilot helps admins, supervisors, and agents work more efficiently through natural language. We cover the migration from Bedrock Inline Agents to Strands Agents, multi-agent orchestration with agents-as-tools, context management strategies, cost optimization, and the testing discipline required to keep agentic systems stable at scale. The 5 lessons: 1. Pick a framework that scales with you — Bedrock Inline Agents worked for 12–15 tools but became exponentially flakier beyond that. Strands Agents gave sensible defaults and room to grow without pinch points. 2. Separate orchestration from domain logic — Agents-as-tools creates a clean line between the orchestrator and sub-agents. You can pull functionality in and out per persona without destabilizing the system, and domain teams own their sub-agents independently. 3. Manage context aggressively — Long context windows for the orchestrator, stateless sub-agents, summarizing and sliding-window conversation managers, and strict control over what tools return. Every extra token in context degrades quality and increases cost. 4. Make prompt caching non-negotiable — System prompts, tool definitions, and conversation history rarely change between invocations. Enabling prompt caching delivered significant cost reductions with almost no effort. 5. Test relentlessly because prompt drift is invisible — One prompt change is never a breaking change; five accumulated changes are. A dedicated weekly Sentinel role investigates failures, and full test suites run on every single change.

Tipu Qureshi — Senior Principal Engineer at AWS — joins the show fresh from the AWS Summit NYC 2026 announcements to break down how DevOps Agent is changing the way teams handle operations and release management. After 14 years across EC2, Elastic Load Balancing, AWS Support, and Networking, Tipu moved into the Agentic AI organization to build the DevOps Agent and contribute to Agent Core. We explore how the agent investigates incidents autonomously, integrates with your IDE through Kiro and Claude, and validates code changes in sandboxes before they hit production. Key takeaways: • Reactive and proactive — DevOps Agent triggers on alarms and ServiceNow incidents, but Custom Agents now run on schedules to detect anomalies before they become outages. • Context is king — Customers who integrate their Git repos, metrics, and logs get significantly more accurate root causes. Native GitHub/GitHub Enterprise support plus bring-your-own MCP for custom observability. • IDE integration — Kiro powers and Claude plugins give on-call engineers the full agentic loop: investigate, root-cause, fix, and validate without leaving the editor. • Release management — The new readiness review inspects pipeline stages, past deployment failures, and integration tests to catch issues before merge, while sandbox testing validates proposed fixes. • Multi-cloud support — Native Azure integration via IDC with RBAC, plus bring-your-own MCP and A2A for on-premises and other clouds. • Custom agents and skills — Bring domain-specific knowledge (SAP HANA failure modes, proprietary tooling) via skills from GitHub repos or the assets API, with MCP tools for full customization. • A2A bi-directional — DevOps Agent can be engaged by other agents and can reach out to other agents, enabling multi-agent escalation workflows. • Transparency — Every tool call, skill invocation, and reasoning step is captured in a journal visible to customers via API and the operator console. • What's next — Deeper integrations, automated mitigation actions with safety policies, time-bound rules for agent escalation, and script execution coming soon.

Davide de Sio built KiroGraph as a personal side project to stop his AI agent from burning through credits just searching files. It turned into a community-driven, open-source MCP server that gives Kiro (and other AI agents) a semantic map of your codebase — reducing token usage by up to 80%. We dive into the architecture, security, and modules, how everything runs 100% locally, and how the AWS Community shaped the project's roadmap. Key takeaways: • Code graphs vs. grep — Tree-sitter and AST-based graph generation give AI agents a smarter navigation model, eliminating wasteful file searches. • Architecture module — Detects patterns and prevents drift by validating your codebase against its own structural rules. • Security module — Finds exposed secrets and vulnerabilities by tracing the call graph, born from an AWS Summit Milano talk. • Watchman module — Auto-generates Kiro skills from repetitive patterns, building persistent memory for your agent. • 100% local execution — Embeddings run with Nomic and summarization with Gemma 3, no data leaves your machine. • Spec-driven development — Davide built KiroGraph with Kiro itself, using specs to drive the entire development lifecycle. • Portability — Commit the graph to Git and share it across machines and team members. • Community-driven roadmap — CI/CD integration, validation hooks, and container deployment are next.

An honest, no-filter conversation about where developers actually stand with AI today. Warren Parad — CTO at Authress, AWS Community Builder, and host of Adventures in DevOps — brings a contrarian 'LLM realist' perspective grounded in daily use, while Romain nuances with enterprise customer observations and the data behind the hype. Together they explore why 93% of devs feel productive but only 4% of enterprises see results — and what separates those who benefit from those who don't. Key takeaways: • AI is a multiplier, not a magic wand — The DORA 2025 report confirms AI amplifies your existing processes. If those processes are broken, AI makes them worse faster. • Spec-driven development beats instant responses — Long-form spec-based workflows let you disengage and return, avoiding the 'TikTok-ification' of software engineering where you're always context-switching. • Sub-agent opacity is a real problem — When agents delegate to sub-agents, you lose visibility into why decisions were made. Custom agents with explicit permissions and tool access help contain the blast radius. • Greenfield work is where LLMs struggle most — LLMs excel at refactoring and targeted feature changes where engineers already know the implementation. Open-ended new projects lead to scope creep and unfinished work. • Critical thinking erosion is measurable — Microsoft/Carnegie Mellon research shows knowledge workers self-report reduced cognitive effort when using AI. The long-term implications for engineering judgment are concerning. • Governance first, tools second — Enterprises that succeed with AI spend the first month on governance, AI registries, and codifying best practices before enabling tools across teams. • Software development was never the bottleneck — Unless AI solves handoffs, knowledge management, and organizational alignment, faster coding alone won't compress your roadmap.

Anthropic deprecated Sonnet 3.5. Some of Xelix's pipelines migrated smoothly. Others broke — and customers noticed within hours. What separated the two? Evaluation. Paul Solomon and James Price Farr have spent 5+ years building AI systems that process millions of invoices for enterprise customers. In this episode, they share the evaluation-first framework that now saves them every time a model changes, an orchestration layer fails, or an agent picks the wrong tool. Key takeaways: • Evaluation-first, not evaluation-after — Retrofitting evaluation on an agent already in production is painful. Build your eval pipeline before you build the agent. • Monitor tool calls, not just outputs — If the agent isn't selecting the right tools, nothing downstream will be correct. Tool-call monitoring is your leading indicator. • 3 tiers of automation — Not everything needs an agent. Rules-based → single LLM call → agentic system. Pick the simplest tier that solves the problem. • Extended thinking tames token explosion — After migrating to newer, more verbose models, enabling extended thinking (with a budget) moved reasoning out of expensive output tokens and brought costs back under control. • Human-in-the-loop by default — Start with human review on every output, then earn trust toward touchless automation as customers gain confidence. • Pragmatism wins — Use whatever technology works best for the problem. Not every feature needs an LLM. Recorded live at AWS Summit London.