Public pages
We built an automated SRE agent on the Agent Development Kit (ADK) and Gemini Enterprise Agent Platform to investigate end-to-end test failures — parsing logs, diffing source code, and classifying each failure as a genuine bug, a PR regression, or a flake. The first iteration produced answers, but suffered from classic agentic fragility: unbounded execution loops despite token guards, non-deterministic routing through identical errors, and confident hallucinations drawn from a single line of log output.
The fix wasn’t a better prompt — it was treating the agent as a software program rather than a series of LLM calls. We encoded the investigation process into the agent’s structure: explicit states, typed transitions, and guardrails defined in ADK, bringing determinism to a highly indeterministic system while cutting prompt volume. This talk walks through the production graph topology, the custom tool boundaries, context engineering under tight platform constraints (no persistent volumes — all source access via GitHub APIs), and the inline verifier that forces the agent to be uncertain and honest rather than confident and wrong.
This talk was presented twice in October 2026, with the Sydney deck significantly revised from the Melbourne original.