Sweetwater Land & Cattle Co.

An AI operations crew for a ranch that is trying to break

An agent loop that watches a simulated Wyoming cattle ranch while a simulator tries to break it. Five crews, one orchestrator, and a person who answers before anything on the ranch is changed.

Built by Scott Blodgett.

The story →One recorded run, told a few beats at a time on a map of the ranch. Two minutes, no background needed.Under the hood →The same run as a machine: the ranch's services, the loop's steps, the models and the database, and what MCP, LangGraph and LangChain actually connect.The control room →Every event the loop wrote, played back on the operator's own screen.

Recorded on 3 October 2026 with the previous models: a small model on our own computer, then Claude Opus 5. Sweetwater runs Claude Haiku 4.5 and Claude Opus 5.5 today.

The hard parts

Code owns severity

Fixed rules in code grade every alarm. The model writes the work order a person reads; it never decides how bad something is, because a model handed a verdict and asked to justify it fabricates the justification.

CLAUDE.md

The AI proposes, a person approves

A model may propose a change to a ranch record and never makes one. The proposal waits at a LangGraph interrupt, checkpointed in Postgres so it survives a restart, until a person answers. Only then does the write go out, over MCP.

src/agent/gate.py

The cheap path is a cheaper architecture

Code gathers each alarm's evidence into one page, so a model judges a page instead of driving a tool loop. Small models write well and navigate badly; that was measured twice before it became the rule.

CLAUDE.md

Measured, then switched off

The one tool loop in the repo, LangGraph with langchain-mcp-adapters on a small local model, ran 28 times on thin pages and improved none of them. It is built, bounded and tested, and it ships off because the measurement said so.

docs/model-routing.md

Nothing fails quietly

The model key once expired and nobody knew for six weeks, because every failure path was quiet by design. Now a free check proves the key before any paid run, and a rejected key raises an alarm every round until it is fixed.

docs/cookbook.md

The measurements

18 of 18work orders Claude Haiku 4.5 wrote that Claude Opus 5.5, reading the same page, did not need to rewrite (14 November 2026)
$0.15what Haiku cost for that whole round: 18 work orders and the shift report
78¢the recorded run's seven rounds with the small model asked first, against about $1.50 with Opus alone
0 of 28thin pages the tool loop improved, which is why it ships switched off

The architecture

The ranch

Four REST services and an MCP server with 19 tools, deployed elsewhere and frozen. The loop reads it every round and writes to it only after a person says yes. It never reads the ranch's source code: the contract is the wire.

The loop

One Python process, one round at a time, 11 steps: read the map, chaos, read sensors, count herd, apply rules, compare, assign crews, gather facts, AI crews, ask a person and report. The first 7 are plain code and cost nothing.

The models and the ledger

Claude Haiku 4.5 answers first and Claude Opus 5.5 takes over when the first answer is thin, both from one key. Postgres holds the alarms, the approvals waiting on a person, and every event these pages replay.

The engineering notes

The repository is private while its history is cleaned for an open-source release.