The Rust Belt knows wasteful systems when it sees them, and AI engineering is running the most expensive idling engine in tech history.
The Summary
- Writer's research shows you can cut token spend by 39% and cost-per-task by 61% — without touching the foundation model or sacrificing accuracy.
- The culprit: "tokenmaxxing," where developers dump entire error logs and context into prompts instead of engineering actual workflows.
- The fix lives in the orchestration layer (the "AI harness") — the code developers already control, no fine-tuning required.
The Signal
Most AI teams are running their production systems like a guy who fixes his car by describing the entire vehicle to a mechanic every time the check engine light comes on. Writer's research names this pattern "tokenmaxxing": shoving maximum context into every LLM call because it's faster than designing a proper system.
The numbers tell the story. Teams watch per-token prices drop and assume their economics are improving. But tokens-per-task is compounding in the other direction. Every loop iteration in an agentic workflow re-transmits the growing context window. The cost reduction from cheaper tokens gets eaten by the exponential growth in token consumption.
"Your invoice is tokens-per-task times price-per-token, and most teams only watch the second number."
Writer's CTO Waseem AlShikh points to a borrowed reflex from traditional software development: generate code, run it, watch it fail, stuff the error message back into the prompt with more context, retry. This works often enough on coding tasks that teams apply it everywhere. The problem is that production agentic workloads aren't one-shot coding problems.
The actual solution lives in the orchestration layer:
- Prompt engineering: tighter instructions, structured outputs
- RAG optimization: better retrieval, fewer irrelevant chunks
- Tool use design: purposeful function calls instead of context dumping
- Workflow architecture: breaking tasks into smaller, sequential steps
None of this requires model fine-tuning or access to model weights. It's all in the harness, the code that wraps the foundation model. The code developers already own and control. Writer's research demonstrates you can achieve 39% token reduction and 61% cost-per-successful-task improvement by optimizing these components while maintaining quality.
The timing matters because we're past the demo phase. Enterprises aren't running proof-of-concepts anymore. They're deploying agents at scale: customer support, document processing, code generation, data analysis. When you're processing millions of tasks monthly, tokenmaxxing turns into a budget crisis. One enterprise support agent might handle 10,000 tickets per day. If each ticket averages 5 loop iterations with 50,000 token context windows, you're burning through 2.5 billion tokens daily on one workload.
The Implication
If you're building or buying AI agents for production, audit your orchestration layer before you negotiate model pricing. The ROI problem isn't the foundation model cost. It's the engineering laziness wrapped around it. Teams with tight harnesses will run production workloads profitably while tokenmaxxers burn budgets wondering why their per-token discounts didn't save them.
This is the difference between companies that treat AI as a research project and companies that ship it as infrastructure. One group optimizes the model. The other group optimizes the system.