Silent Failures, Not Crashes: What Building an Agentic System in Two Weeks Taught Us

I created this article for the purposes of entering the Google Cloud All Things Agentic Hackathon. Two weeks ago, at Elite Solution, we asked a simple question: why do AI support bots confidently answer with policies their companies never wrote? For a small business — a solar installer, a clinic, a workshop — half the real knowledge lives on the website and half in the owner's head. A bot trained on the website alone will hallucinate the missing half. So we built Foundry: an agent that crawls a business website, drafts a support knowledge base, detects the gaps and contradictions it cannot answer — and then does the one thing most bots refuse to do. It asks the human, and waits. The owner answers once, in their own words, and hits Publish. The answer becomes ACTIVE; the old entry is visibly marked SUPERSEDED and kept for history. The support bot answers only from verified, human-approved entries. No answer in the base means a declared gap — never a guess. Autonomous discovery. Human-controlled policy. The stack Everything runs on Google Cloud in europe-west1: Google ADK orchestrates the pipeline (crawler, gap detector, adversarial personas, evaluator, support bot), Gemini 3.5 Flash via Vertex AI powers extraction and evaluation, Cloud Run hosts the service, Firestore holds the DRAFT → ACTIVE → SUPERSEDED lifecycle, and Cloud Scheduler + Pub/Sub + Cloud Run Jobs drive a nightly loop that re-reads the site and files anything new for the owner's approval. We also deployed Gemma 2 (2b-it) on a dedicated Vertex AI endpoint for question classification. The lesson we didn't expect Our biggest takeaway had nothing to do with prompts or models: in agentic systems, the real risk is silent failure, not crashes. Four separate times during the build, a feature looked functional — green tests, clean logs — and was completely inert in practice. A textarea silently destroyed by a polling refresh. A field dropped between two modules. A publish action that worked in the API but was unreachable from the UI. None of it showed up in tests. All of it showed up the moment a human clicked through the real interface. That changed how we work. Before any feature counts as done, someone walks the actual click path, end to end, like a real user. And when debugging under deadline pressure, we confirm the root cause fully before applying any fix — a blind fix is riskier than a known bug. It also changed the product. Foundry's entire premise is that an agent should surface what it cannot do — gaps, contradictions, capability limits — instead of papering over them. Before launch, six adversarial personas (confused, angry, edge-case, off-topic, prompt injection, contradiction prober) attack the bot while a separate evaluator scores every answer. Refusing correctly is scored as correct behavior. In a sense, we ended up building the tool we needed while building it. What's next Wiring the Gemma classifier into the live pipeline, turning the nightly job's findings into a first-class owner review page, and evolving the adversarial loop into a continuous readiness score a business owner can trust before going live. Watch the 3-minute demo: https://youtu.be/1BciPzTgBIs Live instance: https://foundry-1058322784243.europe-west1.run.app Built by Victor Axenciuc, Elite Solution SRL (Moldova), for the Google Cloud All Things Agentic Hackathon. #AllThingsAgentic