Case study · Agent harness
The harness I built when my framework ran out of road.
Offworld worked, but its first version cost about $100 a day and stopped whenever one provider went down. I built Kiger to make the newsroom affordable and dependable, and it runs Offworld in production today.
The problem
Three walls.
I started Offworld on OpenClaw and ran into three limits. Its security wasn't strong enough for anything long-term. It needed to run on a local machine, and I wanted something hosted in the cloud and always on. And the multi-agent setup ran through Discord, where agents sitting in an open chatroom burned tokens just listening. When Anthropic's API went down, everything stopped, and all the context was stuck with it.
The approach
Match the model to the judgment the step needs.
The name comes from the Kiger Mustang herd and Kiger Gorge on Steens Mountain, in Oregon's high desert. Two ideas carry it.
Contextual model switching
Each step runs on the least expensive model that can handle it, chosen by how hard the task is, with a manual override. Different agents can run different models.
Model-agnostic hot-swapping
When a provider goes down, Kiger switches to another model and keeps the context, so the work keeps going.
No secrets in the harness
Credentials live in Owyhee. Kiger runs the agents and picks the models, and Owyhee decides what each agent can touch, every time it reaches for a tool.
The proof
The cost ladder.
Offworld started as an open newsroom: five agents on Opus in one shared room, talking through pitches and drafts. It worked. The agents caught each other's made-up facts, errors, and bad logic. It also cost $100 a day. I wanted to keep the error-catching and stop paying top prices for every step, and each step down the ladder came from the same idea: when a task needs less judgment, it can run on a cheaper model.
| Step | Change | Daily cost |
|---|---|---|
| 1 | Five agents in an open room, all on Opus | $100 |
| 2 | Two-role loop: the journalist pitches, the editor challenges and assigns, the journalist researches and drafts, the editor critiques, fixes go back | $50 |
| 3 | The narrower, more mechanical steps move from Opus to Sonnet | $15 |
| 4 | Writing moves to an open-weights model; a deterministic script checks facts and source links, which the open model sometimes dropped or got wrong | $0.39 |
The result
39 cents a day at the same quality bar.
The biggest savings came from understanding the work well enough to structure it differently: keep the judgment in a few places, give each step the cheapest model that can do it, and use plain code wherever a check is mechanical. Kiger runs Offworld today, and it's where I test everything else I build.
Part of the Kipple Labs stack
Offworld is the publication. Kiger runs its agents and picks their models. Owyhee holds their credentials and enforces what each one may do. AXIS gives them an identity any company can verify.