The latest articles on DEV Community by OJ (@oj-iro). https://dev.to/oj-iro https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4092721%2Fe5a4a742-c5d1-4ed5-9e01-9d763cf17d3b.jpg https://dev.to/oj-iro en OJ Mon, 31 Aug 2026 23:30:32 +0000 https://dev.to/oj-iro/built-an-agent-fleet-that-helps-farmers-for-free-ebj https://dev.to/oj-iro/built-an-agent-fleet-that-helps-farmers-for-free-ebj <p><em>I built this for the All Things Agentic Hackathon. This post covers how it works and what broke along the way.</em></p> <p>A satellite can't see crop disease. It sees one signal — plants looking less healthy than they should — and that has a dozen causes: drought, harvest, haze, a field boundary drawn over a car park. Every satellite tool alerts on that drop, and because most alerts are nothing, farmers stop opening them.</p> <p>So I built <strong>Fieldhand</strong>: a fleet of agents where most of the effort goes into trying to prove its own findings wrong.</p> <h2> The shape of it </h2> <p>Nobody logs in. Cloud Scheduler → Pub/Sub wakes it at 4am, a <strong>Watchman</strong> picks the fields worth looking at, and each runs through a Google ADK fleet: <strong>Diagnostician</strong> (reads the satellite evidence) → <strong>Skeptic</strong> (attacks the diagnosis) → <strong>Agronomist</strong> (plans treatment + a real spray window) → <strong>Operations</strong> (work orders). Anything that costs money stops at a human.</p> <h2> The Skeptic is the whole project </h2> <p>It's not told to "verify" — a model asked to verify goes and finds supporting evidence. It's told to find the reason the diagnosis is <em>wrong</em>. Its main instrument is arithmetic: compare this field to its neighbours, and to its own history in prior years. On a real Palouse field reading −5.5σ (catastrophic), the Skeptic pulled 3 years of history, saw it's harvested and bare every August, and refused to alarm. During testing it also caught a "dying 400-acre field" I'd drawn over downtown Pullman — it looked at the imagery, saw rooftops, and dropped confidence to 0.10.</p> <h2> Three things that bit me </h2> <ol> <li> <strong>On Gemini 3.x, <code>max_output_tokens</code> covers thinking AND the answer.</strong> Budgets ported from another provider got eaten by reasoning and returned truncated fragments with no error.</li> <li> <strong>Gemini sends numeric tool args as strings</strong> (<code>"$6,044"</code>), which crashed the policy layer until I coerced at the tool boundary.</li> <li> <strong>A trailing newline in a secret costs an afternoon</strong> — Python stripped it, TypeScript didn't, every cross-service write 401'd.</li> </ol> <h2> Where the autonomy stops </h2> <p>The policy layer is plain code with no model in it. Even a wrong or prompt-injected agent physically can't spend a grower's money. The interesting engineering wasn't making the agents smart — it was deciding where they <em>don't</em> get a vote.</p> <p>Built with Google ADK · Gemini 3.5/3.7 Flash on Vertex AI · Gemma · Cloud Run · Firestore · Cloud Scheduler + Pub/Sub. Data from Sentinel-2 and Open-Meteo, both free.</p> <p>Live: <a href="https://fieldhand-web-37zk72cqha-uc.a.run.app" rel="noopener noreferrer">https://fieldhand-web-37zk72cqha-uc.a.run.app</a><br> Code: <a href="https://github.com/OJ-IRO/fieldhand" rel="noopener noreferrer">https://github.com/OJ-IRO/fieldhand</a></p> agents ai automation