Talk
Structure Over Prompts: Building an SRE Agent on Gemini Enterprise Agent Platform
We built an automated SRE agent using the Agent Development Kit (ADK) and Gemini Enterprise Agent Platform to investigate end-to-end test failures. The agent's mission: collect and parse logs, analyse source code diffs, locate the root cause, and categorise the failure as a legitimate bug, a PR regression or a flake. Our first iteration produced answers but suffered from typical agentic fragility: edge cases unbounded execution loops even despite token guards, non-deterministic routing paths through identical errors, and overconfident hallucinations such as claiming an "infrastructure failure" based on a single line of a log file. The fix wasn't a better prompt - it was treating the agent as a software program rather than a series of LLM calls. Instead of dumping files, logs, and prompts into an LLM and hoping for the best, we encoded the investigation process into the agent's structure: explicit states, typed transitions, and guardrails defined in ADK. This is what brought determinism into a highly indeterministic system, while reducing the prompts volume. The talk also covers context engineering under time and platform constraints: no persistent volumes or local git clones, so all source code access runs through GitHub APIs. But PR diffs and fetched files can be large, naively adding them to the context will lead to token waste and dilute the attention mechanism, causing the LLM's output quality to degrade. We will walk through the production graph topology, the custom tool boundaries, and the inline verifier that forces the agent to be uncertain and honest rather than confident and wrong.
