Andre Elizondo

AI Summary

→ WHAT IT COVERS Andre Elizondo, head of product at Mesmo, explains how their open-source agent framework Aura applies Kubernetes-style declarative configuration to SRE and platform engineering workflows, addressing why production operations agents require fundamentally different architecture than coding agents across context engineering, multi-agent orchestration, memory, and governed autonomy. → KEY INSIGHTS - **Declarative Agent Configuration via TOML:** Aura uses a simple TOML file to define agents, orchestrators, and workers without scripting reasoning loops or retry logic. Teams specify system prompts, tool access, and LLM selection per agent. Sane defaults handle orchestration mechanics automatically, reducing time-to-outcome from days of framework assembly to same-day deployment. - **Mixed-Model Orchestration for Token Economics:** Assign a large model like Claude Opus to the top-level orchestrator for planning and reasoning, then configure smaller models like Haiku to individual worker agents for delegated tasks. This tiered approach is declared directly in the TOML config and can meaningfully reduce token costs at production scale without sacrificing output quality. - **Scratchpad Pattern for Context Window Management:** SRE workflows generate data volumes that immediately overflow context windows when fed raw into models. Aura's native Scratchpad feature intercepts tool outputs, writes them to disk, and lets the model interact with file contents selectively. Enabling it requires a single config line, preventing context bloat from tools like Prometheus MCP servers. - **Graduated Autonomy via Three Control Layers:** Teams configure human-in-the-loop across three distinct levels: restricting write-access tools entirely, capping how far an agent progresses before requesting approval, or enabling full autonomous remediation. Mesmo's MCP server exposes a single help-request tool that handles on-call lookup and Slack notification automatically, removing the need for agents to resolve those dependencies themselves. - **Runbook Feedback Loop via Agent-Opened PRs:** Rather than treating runbooks as read-only grounding documents, Mesmo configures Aura to open GitHub pull requests updating runbooks when an investigation reveals procedural drift. This closes the institutional knowledge loop, converting one-directional consumption into a continuously maintained, agent-verified source of truth without requiring manual SRE maintenance overhead. → NOTABLE MOMENT Elizondo reframes the SRE role entirely: rather than agents replacing on-call engineers, the outcome Mesmo observes internally is that SREs shift from reactive firefighting toward designing reliability systems, while agents handle tribal knowledge that previously lived only in individual engineers' heads. 💼 SPONSORS None detected 🏷️ AI Agents, Site Reliability Engineering, Open Source Frameworks, Production Observability, Multi-Agent Orchestration