Speak at The Fifth Elephant 2026 Annual Conference

Session Description

Every team building AI agents hits the same wall: the demo works beautifully, then Customer A connects Datadog, Customer B uses Splunk, Customer C has New Relic and a homegrown wiki — and your single agent codebase has to work across all of them without hardcoding anything. This talk walks through the architecture we built to solve this: five layers from user interface to platform, with Skills (composable, tool-agnostic domain logic), Native Tools (platform-owned capabilities that guarantee a Day-1 baseline), and MCP (the protocol that lets the agent discover and invoke tools it has never seen at build time). We cover the three-hop integration gateway, how tenant isolation and delegated auth actually work in production, and why “graceful degradation” isn’t a nice-to-have — it’s the core design requirement.

The second half is the honest part. We shipped this to real enterprise tenants and learned that tool discovery adds 200–800ms per turn (we built a per-tenant cache), that third-party MCP servers return wildly inconsistent output (we built a normalization layer), that OAuth tokens expire mid-investigation (the context ledger enables resumability), and that LLMs treat “no data” as an error when it’s actually valid evidence. We share the architectural decisions that survived contact with production, the ones we had to change, and the composition patterns (skills calling sub-skills, evidence-only synthesis via a context ledger) that eliminated hallucinated data from agent output entirely.

Key Takeaways

1. The architecture that handles partial failure gracefully is the same architecture that scales to new vendors for free — graceful degradation and extensibility are the same design decision, not competing priorities.

2. Evidence-only synthesis via a persistent context ledger (where the agent can only cite what it actually retrieved, never its parametric memory) eliminates the most dangerous failure mode in production agents: confident answers built on invented data.

Target Audience

Engineers and engineering leaders building or scaling AI agents for multi-tenant enterprise environments. Particularly valuable for teams working with MCP integrations, multi-tool orchestration, or any agentic system where tenant isolation, heterogeneous tool landscapes, and partial failure are real constraints — not theoretical ones.

Bio

Venkata sai Varada is a Machine Learning Engineer at Atlassian, where they work on AI solutions in ITOperations domain. He builds AI native features for Incident mnagement, alert managements contributing to AIOps in Jira Service Management.

https://jazzy-dango-501623.netlify.app/

{Add the link to 2-min elevator pitch video}