---
title: "Model vs. Harness: Where AI Agent Controls Belong"
description: "A comparison of what AI models should learn versus what agent infrastructure must enforce. See how reasoning, permissions, isolation, budgets, and observability split across the boundary."
---

# Model vs. Harness: Where AI Agent Controls Belong

**Type:** [Comparison](https://yomu.fyi/visual-explainers/category/comparison.md)

![Two-column comparison diagram titled 'Model vs. Harness'. Left column lists model capabilities: reasoning, coding, tool use, memory, in-context learning. Right column lists harness enforcement: identity, permissions, isolation, budgets, observability, guardrails. A central boundary separates learnable from enforce.](https://yomufyi.s3.us-east-1.amazonaws.com/visual-explainers/9a7572ae-8b75-4ce6-af15-925922b850c8/image-0c0a46a0-feb3-4098-a6be-0a18d1fd1fe0.webp)

> A comparison of what AI models should learn versus what agent infrastructure must enforce. See how reasoning, permissions, isolation, budgets, and observability split across the boundary.

## Explanation

As AI models improve, teams face a recurring decision: absorb a capability into the model, or enforce it in the surrounding harness? This explainer compares model-side capabilities such as reasoning, coding, tool use, and memory with harness-side enforcement such as identity, permissions, isolation, budgets, and observability. The key condition is that better models make weak harnesses more dangerous, so enforcement must stay outside the model even as capabilities move in. Engineers can use this boundary to diagnose where an agent's controls belong and avoid removing scaffolding too early.

## FAQ

### How do I decide what to absorb into the model versus keep in the harness?

Absorb capabilities the model can reliably learn, such as planning, tool use, and reasoning. Keep enforcement properties—permissions, identity, isolation, and budgets—as hardened infrastructure outside the model. The harness engineering post warns that removing enforcement scaffolding as models improve is dangerous.

### Does a stronger model make a weak harness safer or more dangerous?

More dangerous. The better the model gets, the more it can chain vulnerabilities across trust boundaries when the harness gives it paths around controls. A breach at Hugging Face showed this pattern, so enforcement must remain outside the model.

### What practical layers enforce controls in an enterprise AI agent?

Use an API gateway for identity and policy, sandboxed containers for isolation, a gateway layer for model routing and budgets, and distributed tracing for observability. These map to zero trust, defense in depth, and fail-safe defaults from the governed agent reference architecture.

**Tags:** [Agentic AI](https://yomu.fyi/topic/agentic-ai.md), [AI Agents](https://yomu.fyi/topic/ai-agents.md), [AI/ML](https://yomu.fyi/topic/ai-ml.md), [Guardrails](https://yomu.fyi/topic/guardrails.md), [Prompt Injection](https://yomu.fyi/topic/prompt-injection.md), [Sandboxing](https://yomu.fyi/topic/sandboxing.md)

## Related reading

### [Introducing Claude 4](https://yomu.fyi/post/introducing-claude-4.md)
- Company: [Anthropic](https://yomu.fyi/company/anthropic.md)
- Author: Anthropic
- Published: Nov 12, 2026
- Type: Announcement

Anthropic introduced Claude Opus 4 and Claude Sonnet 4, setting new standards for coding, advanced reasoning, and AI agents. Claude Opus 4 is described as the world’s best coding model, leading on SWE-bench (72.5%) and Terminal-bench (43.2%), with sustained performance on long-running tasks. Claude Sonnet 4 significantly improves on Sonnet 3.7, achieving 72.7% on SWE-bench and excelling in agentic scenarios. Both models are hybrid, offering near-instant responses and extended thinking, which now supports tool use, parallel tool execution, and improved memory via local file access. Claude Code became generally available with VS Code and JetBrains extensions, GitHub Actions background tasks, and an extensible SDK.


### [Inside an enterprise AI agent: A governed, observable loop architecture](https://yomu.fyi/post/inside-an-enterprise-ai-agent-a-governed-observable-loop-architecture.md)
- Company: [AWS](https://yomu.fyi/company/aws.md)
- Author: Rushabh Lokhande
- Published: Nov 12, 2026
- Type: Explainer

An enterprise AI agent is not a larger chatbot; it is a multi-step system that plans, retrieves knowledge, calls tools, and verifies results in a loop. This reference architecture organizes the agent into five layers: an API gateway for identity and policy, a five-phase agent runtime (Understand, Plan, Retrieve, Reason, Act and Observe), a knowledge and state tier, a platform-services layer, and a cross-cutting trust and control framework. Each layer maps to AWS managed services such as Amazon Bedrock, AWS Lambda, and Amazon EventBridge. The design follows the AWS Well-Architected Framework Generative AI Lens, applying zero trust, defense in depth, separation of concerns, observable by default, and fail-safe defaults. Governance, observability, and cost controls must be embedded in every layer, not bolted on at the end.


### [LLMON: An LLM-Native Markup Language, now in Granite 4.1 3B and 8B](https://yomu.fyi/post/llmon-an-llm-native-markup-language-now-in-granite-4-1-3b-and-8b.md)
- Company: [Hugging Face](https://yomu.fyi/company/hugging-face.md)
- Author: Abraham Daniels, Basel Shbita, Chad DeLuca, Dan Gutfreund, Michael Hind, Shub A
- Published: Nov 12, 2026
- Type: Announcement

LLMON is a markup language that uses reserved special tokens to explicitly communicate structure and semantics to LLMs, addressing brittleness and prompt injection vulnerabilities. Components like instructions and data are marked with meaningful tags and unique IDs, enabling explicit binding between them. The authors release two experimental preview models derived from Granite 4.1 base models, trained to consume and generate LLMON-marked text. On the OPI benchmark, LLMON models achieve near-zero attack success rates for indirect prompt injections, outperforming the best frontier model by 47% while maintaining general capability comparable to Granite 4.1 Instruct models. LLMON enriches chat templates by adding component-level semantics and binding, and the models are trained through re-serialized instruction-tuning data and reinforcement learning.


### [How Block powers collaborative AI agents with Buzz and Unity Gateway](https://yomu.fyi/post/how-block-powers-collaborative-ai-agents-with-buzz-and-unity-gateway.md)
- Company: [Databricks](https://yomu.fyi/company/databricks.md)
- Author: Bradley Axen, Brandon Kvarda, Julia Powell, Kelly Albano
- Published: Nov 12, 2026
- Type: Announcement

Block developed Buzz, an open source collaboration workspace where employees and AI agents work together in channels, threads, and direct messages. Users create agents, choose a model or coding harness per agent, and collaborate with multiple agents and colleagues in one environment. Databricks Unity Gateway provides centralized access to frontier and open-weight models from OpenAI, Anthropic, and Google through a single integration. Block experiments with multi-agent workflows, using a more capable model to coordinate agents powered by smaller, more efficient models, and has an internal benchmarking approach for different configurations. Buzz is model-agnostic, self-hostable, and supports multiple agent harnesses including a lightweight version of Goose.


### [Genie in Slack is now Generally Available: Governed AI cowork, right where work happens](https://yomu.fyi/post/genie-in-slack-is-now-generally-available-governed-ai-cowork-right-whe.md)
- Company: [Databricks](https://yomu.fyi/company/databricks.md)
- Author: Toussaint Webb, Sydney Sundell
- Published: Nov 12, 2026
- Type: Announcement

The Databricks Genie app for Slack is now generally available, allowing teams to ask data questions and receive governed answers directly in Slack channels, threads, or DMs. Genie uses the asker’s Unity Catalog and source permissions, so it only surfaces information that person is allowed to see. Every answer cites its sources, and users can click through to explore deeper in Genie One. Beyond answering questions, Genie can take action via MCP connectors, schedule recurring analyses, and work with uploaded files such as PDFs. Admins govern AI usage centrally in Unity Gateway, including MCP policies, cost controls, and monitoring, and Private Link workspaces are supported through context-based ingress policies.


### [Sandbox resourcing: burst, don't reserve](https://yomu.fyi/post/sandbox-resourcing-burst-don-t-reserve.md)
- Company: [Modal](https://yomu.fyi/company/modal.md)
- Author: Kevin Yang
- Published: Nov 12, 2026
- Type: Explainer

Modal Sandboxes are bursty workloads that spend most of their time using far less than peak CPU, making reservation-based pricing wasteful. The post argues that reserving capacity for each Sandbox’s peak would require machine capacity equal to the sum of all peaks, while bursting allows multiple Sandboxes to share capacity efficiently. Modal uses active-resource pricing where Sandboxes are billed for a default request when idle and pay for additional resources only when they burst. CPU can burst up to 16 cores above the request, and memory has no limit by default, though limits can be set for consistency or cost capping. Evictions are rare, with an SLO of under 0.0001% OOM rate per minute for Sandboxes under 8 GiB, and tools like exit snapshots and volumes help mitigate their impact.


### [Agent harness engineering: Separating model capabilities from enforcement](https://yomu.fyi/post/agent-harness-engineering-separating-model-capabilities-from-enforceme.md)
- Company: [Red Hat](https://yomu.fyi/company/red-hat.md)
- Author: Mark Vaykhansky, Uri Shaket
- Published: Nov 12, 2026
- Type: Explainer

A breach at Hugging Face showed that capable agents can chain vulnerabilities across trust boundaries when the surrounding harness gives them paths around security controls. The post argues that capabilities models can reliably learn should be absorbed into the model, while enforcement properties such as permissions, identity, isolation, and budgets must be hardened as infrastructure outside the model. It classifies harnesses into four types—workflow, coding, research, and agent runtime—each with different architectures for different jobs. The better the model gets, the more dangerous a weak harness becomes, because teams often remove enforcement scaffolding as the model absorbs planning and tool use. The post recommends diagnosing gaps in an existing agent’s harness before applying tooling such as a gateway layer, sandboxed containers, or distributed tracing.
