One application. 4B at the edge. Ten agents at the frontier.
-
Not two model integrations. One TypeScript harness spanning both.
-
Lloyal brings live model execution into the application itself, letting one intelligent program contract to edge-scale weights or expand into a frontier agent population.
+
+
+
Category / Application Architecture
+
Vertical Inference
+
System intelligence beyond the model.
+
+
+
The application, model runtime and live inference state become one programmable system.
+
Vertical Inference is an application architecture in which model execution becomes resident application state. Ordinary code governs how intelligence retrieves, reasons, acts and continues while it is underway.
+
Lloyal is the platform for building and deploying this architecture—from local applications to frontier compute.
The weights and data move inside your infrastructure. The product still receives tokens through an endpoint.
-
-
-
LLOYAL · HARNESSED MODEL
-
Run the intelligent application
-
-Product surfaces↓
-One intelligent application harness · policy · capabilities · branch topology↓
-One live model-state population↓Resident model execution
-
-
Model execution and product behaviour run as one application across CLI, desktop and web.
-
-
-
-Same weights. Same infrastructure.
-Bare models expose tokens. Harnessed models carry the application.
-
-
-
-
02 / THE PROGRAMMING SURFACE
-
MVC, with a live LLM as the Model.
-
Your product surfaces are the Views. Your harness is the Controller—ordinary TypeScript defining how the application's intelligence collaborates, acts, recovers and continues. Beneath it, a resident LLM holds the live generative state.
+
01 / THE HARNESS
+
The harness is the intelligent application.
+
You already know how to build this: MVC, with a live LLM as the Model. Product surfaces are the Views. The harness is the Controller—ordinary TypeScript defining how the application's intelligence collaborates, acts, recovers and continues. Beneath it, a resident LLM holds the live generative state.
Views
CLI · Desktop · Webproduct surfaces
↓
@@ -148,7 +754,7 @@
MVC, with a live LLM as the Model.
↓
Model
Resident LLMshared context · live branches · generative state
-
Program intelligent behaviour—not just prompts and calls.
+
The model supplies the intelligence. The harness supplies the institution.
Start with the platform. Or build an intelligent product capability with us.
-
Use the shipped developer tooling yourself, follow the deployment path as it lands, or work directly with Lloyal to make a high-value capability part of the product your customers already use.
-
-
-BUILD
Start with the developer platform.
Scaffold a harness, inspect the runtime and extend the application with signed AgentApps.
Create the intelligent application. Wire the product around it.
-
-
harness.dev new scaffolds the headless application today. The same program can then project into CLI, Desktop and Web without re-authoring its intelligence.
-
-
-
AVAILABLE NOWdeveloper entry point
-
Start with the application program.
-
-$npx harness.dev new claims-review
-
-
Scaffolds a TypeScript harness from a basic template — along with a CLI frontend and AgentApps.
Each interface projects the same headless entrypoint. The shell changes; the intelligent application does not.
-
-
-
+
+
02 / AGENTAPPS
-BUILT FOR MODELS YOU CAN RUN
-
Run the weights on-device, on a workstation, in your cloud account or inside customer infrastructure.
+
Intelligence needs contact with reality.
+
AgentApps connect the harness to documents, browsers, feeds, enterprise systems and tools. They can retrieve evidence, observe changing state and perform authorised actions. The harness determines when they run, which live lineage receives what they return and what consequence is permitted.
-
-Apache harness.dev CLI
-FSL runtime
-Reviewed & signed AgentApps
+
+
Retrieve
Acquire and evaluate evidence across search, corpora, documents and databases.
+
Observe
Read the world as it changes through browsers, feeds, telemetry and application state.
+
Act
Produce permitted consequence through computer use, workflows, APIs and operational tools.
+
+
AgentApps are signed, portable capability contracts combining Skills · Live Sources · Tools · Grants · Protocol
+AgentApps provide contact with reality. The harness supplies procedure and authority.
-
-
-
-
-
05 / COMPUTE
-
One programming surface. Two radically different compute envelopes.
-
-
-
-
01
-
SCALE DOWN
-
The harness amplifies the model.
-
A quantised 4B model runs planning, human review, multi-agent investigation, tools, pressure-aware recovery and synthesis—not a local chatbot.
-
-
Model
Qwen 3.5 4B · quantised
-
Directional RACE
0.405
-
Reference anchor
Claude 3.7 · 0.422
-
-
Directional English subset — the 4B scored 4 of the benchmark's 5 English tasks (task 83 planner-stalled); not a leaderboard submission or same-task Claude comparison. Citation grounding was the identified weakness.
Ten agents. One model dispatch per step. One GLM-5.2 deployment advanced ten research agents over a shared decoded context as one live branch population.
-
-
Concurrent agents
10
-
Model deployment
1
-
Shared contexts
1
-
Decode dispatches
1 per generation step
-
-
Fan-out still consumes KV capacity and generated tokens. It does not duplicate the resident model or reconstruct the shared prefix for every agent.
+Choose the surfaces, resident model and starting procedure. The scaffold installs a working application you can change.
+
@@ -284,9 +825,9 @@
The harness compounds the model.
-
DESIGN PARTNERS
-
Build the part of your product that thinks.
-
Work with us to turn one high-value product capability into a harnessed intelligent application—integrated into the existing product and deployable across the environments its customers require.
+
FOR PRODUCT TEAMS
+
Make one high-value capability native to your product.
+
Work with Lloyal to turn a consequential workflow into a production harness—integrated with your existing product, connected to the systems it depends on and deployable across the environments your customers require.
One product capabilityOne product integrationPrivate AgentAppsMultiple placements6–10 weeks
+