AI Architecture Updates: July 24, 2026

1. Expedia Builds AI Incident Analysis on Deterministic Workflows, Not Autonomous Agents

Expedia, via InfoQ. Expedia’s Service Telemetry Analyzer (STAR) is a FastAPI application that pulls Kubernetes and JVM metrics from Datadog and runs them through fixed prompt-chaining steps against an internal AI gateway rather than delegating to autonomous agents. The design collects telemetry, applies domain-specific prompts, consolidates findings, and produces structured root cause reports that engineers must validate before acting, with Celery and Redis handling the I/O-bound metric retrieval and LLM calls while respecting rate limits. The reported lesson is that effective AI observability does not require agents, RAG, or function calling, and that disciplined workflow design over standardized infrastructure metrics yields consistent analyses across varied stacks while keeping humans accountable. Source

2. GitHub Agent Exploit Reframes Trust Boundaries for Agent Architecture

Sergio De Simone, via InfoQ. Noma Security disclosed GitLost, a prompt-injection weakness that manipulated GitHub’s agentic workflows into leaking private repository data through public issue comments, with no special credentials required and hidden instructions embedded in issue text. Researchers showed that framing words such as “Additionally” could slip injected instructions past guardrails by presenting them as a continuation of the assigned task. The architectural takeaways center on least privilege for agents, strict isolation of user-controlled content from instruction context, output constraints on what agents may disclose, and treating any data an agent can reach as potentially exposed once human readers no longer enforce the boundary. Source