Build an SRE Agent Harness for AIOps Without Context Blowout

Mezmo

Jul 9, 2026

See Aura in action → https://www.mezmo.com/aura

An agent harness for AIOps is the runtime layer that coding agents like Claude Code were never built to provide: context isolation, decision traceability, and gated execution for tools that touch production. Aura is Mezmo's open-source (Apache 2.0) agent harness, purpose-built for operations work rather than software development. In this PlatformCon 2026 session, we cover the problems that forced it into existence — a 15-minute log window producing 10 million tokens, and a needle-in-a-haystack investigation that looked correct while silently dropping 16% of its findings — then walk through the orchestrator/worker DAG architecture that isolates context per worker, the intent disclosure that doubles as an agent audit trail, human-in-the-loop webhook gating for destructive tools, and a live root cause diagnosis run against a real loaded environment.

Aura is built for SREs and platform engineers who need autonomous incident investigation without handing production remediation to an unaudited black box.

Chapters:

00:00 Why not just use Claude Code with an MCP server

02:53 What an agent harness actually does

04:09 Building Aura at a telemetry company: 10M tokens in 15 minutes

06:38 Going production-first and fixing edge cases early

07:57 The lossiness trap: 16% of findings silently missing

09:35 Orchestrator and worker DAG architecture

10:41 Intent disclosure as an audit trail

12:21 Setup walkthrough: the repo, the binary, aura init

18:51 Familiar chat mode and standalone web server

20:25 Live demo: diagnosing a real incident from a sloppy paste

24:23 Human-in-the-loop: webhook gating for destructive tools