Reliability in an AI agent is a harness property, not a model property.
The cleanest proof arrived at the bottom of the model-size ladder: a 688 MB model controlling a smart home, showcased by the model's own maker. The part worth studying is the 25 MB of Rust wrapped around it.
EdgeHome Harness pairs MiniCPM5-1B, 688 MB on disk and about 1 GB at runtime, with a Rust harness between 25 and 64 MB, targeting devices with 2 GB of RAM. A community developer built it; OpenBMB put it on stage. The line everyone will quote is that a full agent stack fits on ultra-constrained edge hardware. The design decision that matters is different: the model is not allowed to do anything.
Everything After the JSON Is Deterministic
EdgeHome Harness is a Rust framework that makes a 1B model act as a reliable smart-home controller. The model only generates a candidate JSON command. Deterministic Rust code validates the schema, resolves devices against a registry, checks capabilities, applies fail-closed policy gates, plans a dry run, and records a replayable trace.
The repo states the thesis outright: the model sits at the top of the pipeline, and everything after candidate generation belongs to the harness. There is a trust table in the docs. The model's output type is labeled untrusted. Not partially trusted. Untrusted, the way you treat user input in a web form.
The harness ships payload adapters for Home Assistant, MQTT, and MIoT or Matter bridges. The model itself talks to nothing but Ollama, through a structured output request.
The Model Is the First Thing Sacrificed
A 1B model repeats itself, rambles past the closing brace, and emits JSON that almost parses. EdgeHome budgets for all three: an output governor with dead-loop detection, a retry policy, and fallback classification. Short-term memory stays out of the context window too; Rust tracks the last target and confirmed aliases, because context is the scarcest resource on a 2 GB board.
The sharpest detail is the pressure policy. When memory runs low, the harness shrinks the model's context, then its output budget, and under critical pressure drops the model entirely and falls back to rules. In this architecture, the language model is the most disposable component in the stack. Execution backends ship disabled by default, every plan is a dry run first, and a 108-case eval gate decides releases.
The Same Boundary Holds at Frontier Scale
I run the identical pattern at the opposite end of the size ladder. My coding agent works inside a sandbox where the harness owns the filesystem and network boundaries, whatever the model believes. My wallet experiment puts a policy gate with spend caps and an allowlist between an agent and the chain, because an approval step the agent can call is an approval step the agent can forge. Enforcement lives outside the model or it does not exist.
I have made this argument for coding agents and for cheap-model pipelines. EdgeHome is the same argument compiled to 25 MB and pointed at a light switch.
This project settles a different question than the one in the announcement. Not whether a 1B model can run your smart home. How few decisions a model of any size should be making. Trace tonight's agent failures back to their source: how many were the model being too small, and how many were your harness letting the model decide something a validator should own? EdgeHome's answer is a trust table. Most agent stacks never wrote one.
I write field notes from real builds — AI integration, cron-driven automation, and the parts that break in production. New posts every two weeks; if this one was useful, notes on human-in-the-loop approval boundaries for agents is the companion download.
