Rent the model, own the harness: how to build a model-agnostic AI agent

Werk developed an agent scaffolding system for mobile field teams that separates foundation models from application business logic, according to blog.getwerk.ai's report. While developers have Claude Code or Codex and office workers use Claude Cowork or ChatGPT Work, rooftop installers lack purpose-built tools. Werk built its platform so models propose actions while a custom harness controls data access, permissions, and records across cloud providers.

This report examines the architectural principles behind building model-agnostic agent software for field crews. The analysis details how custom execution loops, dynamic provider configuration, structured tagging, and prompt prefix caching protect operational workflows from vendor outages. It is written for software architects and engineering leaders evaluating custom agent harnesses against generalized frameworks, enterprise authorization boundaries, and multi-tenant security requirements in production environments.

Rent the model, own the harness: what actually changed

Enterprise AI development is shifting away from generalized agent libraries toward hand-written execution loops that treat frontier models as interchangeable utilities. A 2026 arXiv study by Barbaste, Darrigol, Vu, and Wiltberger reviewed the source code of eleven coding harnesses, including Claude Code, Codex, Gemini CLI, Hermes, and Pi. The researchers revealed that none of the eleven harnesses imported a general-purpose agent framework, relying instead on custom loops. This matches Anthropic's 2024 Building effective agents guidance, which advised developers to call model APIs directly rather than using complex frameworks.

Werk implemented this separation after studying more than fifteen open-source projects licensed under MIT or Apache 2.0, including Aider, OpenHands, Codex, Gemini CLI, and Pi. The company applied these insights to field tasks, such as processing a four-second roof voice note stating that the Novak job is done. While the model interprets the audio transcript and identifies the customer card, the harness manages execution. The harness selects visible cards, validates user permissions, enforces idempotent recording across duplicate transmissions, and notifies assigned team members.

Rent the model, own the harness: how it works

The architectural separation between the harness and the model functions through four engineering practices. First, the application accesses models exclusively through an internal interface where providers are managed as runtime configuration. When provider status dashboards from Anthropic or OpenAI report outages, teams can redirect traffic without altering application logic. Second, individual agents declare a required level of work rather than specifying a model name. An automated test battery continuously evaluates candidate models against concrete scenarios to determine which provider fills each operational tier.

Third, the architecture standardizes input formatting across models by adopting prompt guidelines published by Anthropic and OpenAI. The harness isolates conversation turns, job cards, and voice transcripts inside structured semantic tags so every model receives identical context. Fourth, prompt builders place stable instructions first and append dynamic request data last. Upstream model providers reuse previously processed prompt prefixes, accelerating response times while lowering expenses; in DeepSeek Harness benchmarks, this prefix caching strategy achieved a 98 percent cache hit rate across 650 requests.

Rent the model, own the harness: why it matters

Agent SDKs like the OpenAI Agents SDK, Claude Agent SDK, LangGraph, Vercel AI SDK, Pydantic AI, and Mastra handle the basic cycle of prompting models and executing tools, but they cannot resolve multi-user team synchronization. In physical field work, customer records like Novak's card are shared across crew members with distinct access privileges. When an agent asks a worker to choose between two candidate cards and the worker replies with the second one, that confirmation must bind strictly to the visual state displayed on that specific user's screen.

Mastra documentation highlights this necessity by recommending that approval actions bind directly to the exact arguments presented to the human reviewer. Furthermore, single-tenant personal assistants cannot manage shared team records. Open-source agents such as Hermes Agent and OpenClaw assume a single trusted operator; the Hermes security policy explicitly identifies the system as a single-tenant assistant where all users share equal trust. In collaborative team deployments, the harness must enforce role permissions deterministically, ensuring that non-deterministic language models never decide who can access or modify shared project records.

Rent the model, own the harness: who is affected

Mobile trade crews, including solar panel installers, electricians, and field service technicians, represent the frontline workforce affected by this architectural design. Unlike office workers using Claude Cowork or ChatGPT Work, physical tradespeople operate hands-free in challenging environments where software interactions occur through brief audio clips recorded on rooftops. Unstable mobile connectivity and environmental noise frequently cause duplicate message delivery and transcription ambiguities. A dedicated harness prevents duplicate database writes, verifies local worker authority, and protects field workers from interface breakdowns caused by underlying model updates or transient cloud provider service interruptions.

Software developers and system architects building enterprise AI agents are also directly affected by these design principles. The empirical evidence across academic research and commercial products indicates that generalized agent frameworks introduce brittle abstractions that hinder production scalability. By building modular execution harnesses around commodity foundation models, engineering organizations maintain full ownership over access control lists, prompt caching structures, error recovery, and data security boundaries. This decoupling lets engineering teams swap or benchmark new language models immediately as performance and pricing improve without disrupting core business workflows.

Rent the model, own the harness: what to watch

Werk has opened a public waitlist for early access while its field operations software remains in private development. The company is publishing an ongoing technical blog series detailing its architectural framework, with the upcoming installment examining its core dispatch router. Technical observers should watch how this router inspects incoming voice notes and maps user intent to specialized agents across defined work tiers. Evaluating this implementation will illustrate how production teams manage intent classification, latency overhead, and dynamic model routing without relying on external multi-agent framework abstractions in real-time mobile environments.

Broader industry developments to monitor include how model providers adapt prompt prefix caching and enterprise concurrency features. As inference costs continue to decline, structural caching techniques like the 98 percent hit rate recorded in DeepSeek Harness benchmarks will define production unit economics. The primary technical challenge going forward is maintaining rigorous boundaries between non-deterministic model proposals and deterministic application state: foundation models will continue to advance in language comprehension, but operational durability will depend on how reliably independent execution harnesses enforce permissions, data validation, and multi-user transactional consistency.