When we talk about AI coding agents, we usually obsess over the model. Is Claude 3.5 better than GPT-4o? Does the latest Llama release hold the crown? But there is a hidden player in the room that might be impacting your wallet and your workflow more than the LLM itself: the harness.
What Exactly is the 'Harness Tax'?
A harness is the software layer sitting between the AI model and your codebase. It handles the heavy lifting—deciding which files the model sees, managing tool execution, and applying quality gates. The "Harness Tax" is the efficiency gap created by this layer. It is the cost of the fixed context, system prompts, and overhead that the harness resends with every single turn before the model even sees your actual instruction.
Model vs. Harness: Who Wins?
Recent research suggests that the choice of harness can actually dwarf the effect of the prompt. In some tests, the cost per success ran 5x to 30x higher through complex harnesses compared to minimal ones. For example, while a minimal harness like Pi serves as a lightweight reference point, more "managed" agents can introduce significant dead weight.
This becomes a critical issue as sessions grow. When your context window fills up, the way a harness handles caching and token resending determines your run cost more than the model's headline rate. Essentially, you are paying for capacity that doesn't always buy you more signal.
The Future of Agent Efficiency
As we move toward 2026, the industry is realizing that managed agents aren't necessarily a new capability—they are often just a wrapper. The real challenge moving forward isn't just building a smarter model, but building a leaner harness that provides project-scoped memory without the bloated token tax.
Sources
Media
