Performance Comes From the Harness
Two companies, the same AI model, opposite results. The difference is not the model but the harness. The 5 layers explained for a CTO or CISO.
The essentials in 5 points
- Two companies with the same model get opposite results. The difference is not the model, it is the harness.
- The harness is the software layer around the LLM: it gives the model the memory, tools, and limits it lacks on its own.
- A robust harness rests on five layers: identity, memory, tools, triggers, guardrails.
- Adding layers does not make the agent better. Over-harnessing makes the agent as unpredictable as under-harnessing.
- When choosing a provider: some sell access to AI, others build the harness. These are not the same product.
The short answer
When two companies use the same AI model and get radically different results, the cause is almost never the model. It is the harness: the software layer that wraps the model and decides what it sees, what it keeps, what it can do, and what stops it. The model is the brain, the harness is the body.
A robust harness is organized into five layers: identity (the persistent system prompt), memory (what survives context resets), tools (dedicated capabilities rather than a generic chat), triggers (what makes the agent act without being asked), and guardrails (what bounds its autonomy). The rest of this article details each one, the design trade-offs, and what changes when you choose a provider.
The model has become the easy part
There is a stubborn reflex in technical leadership: to believe that the quality of an AI deployment depends first on the choice of model. That was true two years ago. It no longer is. Frontier models are converging in capability and commoditizing in access. A careful reading of mature agentic systems leads to a clear conclusion: you could swap the underlying model for another and, with a little optimization, recover a comparable level of performance. The difference lies in the orchestration, not in the model itself.
In other words, differentiation has left the raw model and moved toward memory and context. A tool like Claude Code is not first a terminal chat, it is an agent orchestration platform that presents itself as a chat. The value is in the harness. For a CTO, that shifts where attention and budget should go: less on picking between two models, more on the engineering that surrounds them.
The five layers of a harness
1. Identity. This is the persistent system prompt, re-read on every turn rather than once at startup. In the Claude Code ecosystem, it is the CLAUDE.md file: a few thousand characters that define the agent’s behavior, its working principles, its conventions. It is the highest-leverage configuration point in the whole system, because it changes the agent’s behavior without ever touching the model.
2. Memory. A model forgets everything between sessions and degrades as its context grows. Memory is what survives those resets: files read at startup and rewritten at the end of a session, a database for structured queries, a consolidated snapshot passed from one session to the next. Without this layer, the agent starts every task from scratch and keeps nothing from the previous one.
3. Tools. A useful agent is not a model that “chats with files,” it is a model equipped with dedicated capabilities: skills that codify domain expertise, connectors to internal systems, precise commands. A well-designed tool does more than expose data, it encapsulates a reliable procedure that the agent triggers at the right moment.
4. Triggers. A mature harness does not only answer when spoken to. Hooks and scheduled tasks make it act on its own: checking a repository at regular intervals, reacting to an event, restarting a loop. This is what turns the agent from an assistant you prompt into a system that works continuously.
5. Guardrails. An agent can act on its own because protections frame it, not because they have been removed. Granular permissions, isolation of critical components, deterministic blocking of destructive actions, human review on sensitive operations (production, migrations, shared infrastructure). This layer does not depend on the model’s good will: these are coded rules that apply whatever the model produces.
The design trade-offs
Laying down the five layers is not enough, they have to be tuned. Three trade-offs recur in every deployment.
Loop or orchestration? A classic agentic loop lets the model decide when to stop, which suits exploratory tasks. An orchestrated plan that splits work across several agents suits broad, reproducible tasks. The real skill is no longer to phrase a better instruction, it is to architect the right loop.
Over-harnessing or under-harnessing? Too many skills, rules, and triggers bloat the context and make behavior unpredictable. Too few, and the agent wanders. The right balance is a short, behavioral identity prompt, not a wall of rules. Adding a layer has a cost, and it has to be justified.
Fixed or reassessed? A harness is never final. Every new model release moves the line of what still needs scaffolding. A stronger model makes unnecessary the structures you had built to compensate for its weaknesses. The discipline is to reassess the harness with each release and strip out the scaffolding that has become redundant.
What changes when you choose a provider
In the market, two offers coexist, often described with the same words. The first sells access to AI: a model, an API key, a chat interface, sometimes a convincing demo. The second builds the harness: the domain identity, the memory that compounds, the tools that encapsulate your expertise, the triggers that keep the system running, the guardrails that keep it under control.
The first delivers something that impresses in a meeting and disappoints in production, because nothing keeps a memory of the work, nothing frames the agent, and nothing tracks how models evolve. The second delivers a system you can genuinely hand work to. The question to ask a provider is therefore not “which model do you use?” but “how do you manage memory, where do you draw the line between automatic and human, and how does your harness evolve when the model changes?”. The answers to those three questions tell a demo apart from a real deployment.
At Colombani.ai, this five-layer architecture is the backbone of every business AI agent deployment and the core of the Claude Code in Production course, delivered by a Qualiopi-certified provider. The expertise brought to bear is certified by Anthropic (Claude Certified Architect), and the harness delivered is designed to stay under your control and evolve with every model release. The harness is also judged tool by tool: the 2026 CLI agents comparison measures cost, context, MCP and memory for Claude Code, Gemini CLI, Codex, Hermes and OpenClaw.
Sources
Frequently asked questions
What is an AI agent's harness?
The harness is the software layer wrapped around the language model that turns it into a reliable system: memory management, tool access, automatic triggers, guardrails, and a persistent system prompt. The same model placed in two different harnesses produces radically different results, because the harness decides what the agent sees, keeps, can do, and cannot do.
Why do two companies with the same model get different results?
Because an agent's performance does not come from the raw model but from the engineering around it. Two companies using the same base model, one with plain conversational access and the other with a full harness (persistent memory, dedicated tools, deterministic guardrails), do not have the same product. When models commoditize, differentiation moves toward orchestration, memory, and context, which is to say toward the harness.
What are the layers of a robust harness?
Five layers: identity (the persistent system prompt, re-read every turn, that defines the agent's behavior), memory (files and databases that survive context resets), tools (codified skills, connectors, dedicated commands rather than a generic chat), triggers (hooks and scheduled tasks that make the agent act without human intervention), and guardrails (permissions, isolation, deterministic blocking of dangerous actions).
Should you add as many layers as possible to an agent?
No. Over-harnessing degrades reliability as much as under-harnessing: too many rules, skills, and triggers make behavior unpredictable and bloat the context. The right setting is a short, behavioral system prompt, not a wall of rules. The key rule is to reassess the harness with every new model release and strip out the scaffolding that a stronger model has made unnecessary.
How do you choose a provider to deploy AI agents?
The dividing line is simple: some sell access to AI (a model, an API key, a chat), others build the harness that makes the agent reliable and able to work on its own. The first delivers a demo, the second a system that builds on experience, stays under control, and evolves with every release. A serious provider talks about identity, memory, guardrails, and what the agent is allowed to do on its own, not just about model quality.