TL;DR
- The software layer orchestrating your AI agents — not the model itself — can account for up to 38% variance in token consumption on identical tasks.
- Most enterprises have no visibility into harness-level costs because configuration is fragmented across multiple teams with no single owner.
- FinOps for AI agents is roughly where cloud FinOps was in 2013 — a discipline that barely exists yet.
- As agentic AI scales from pilot to production, the harness is quickly becoming the most consequential — and most ignored — lever in enterprise AI economics.
The Question Enterprises Are Finally Asking
For the past two years, boardrooms and IT procurement meetings have been consumed by one question: which large language model should we use? GPT-4o or Claude? Gemini or Llama? The debate has been fierce, well-funded, and, it turns out, slightly beside the point.
A growing body of research and real-world testing is pointing to an uncomfortable truth: once you've picked your model, you may have already made the least impactful cost decision in your entire agentic AI stack. The layer that actually determines whether your enterprise AI burns money efficiently or inefficiently isn't the model. It's the agent harness — the orchestration software that coordinates your model, tools, workflows, and context. And right now, most enterprises are flying completely blind on it.
What Is an Agent Harness, and Why Should You Care?
If the LLM is the engine of an AI agent, the harness is the chassis, transmission, fuel injection system, and GPS navigation — all rolled into one invisible layer. It handles system prompts, tool definitions, agent coordination, Model Context Protocol (MCP) server connections, subagent fan-out, and a dozen other orchestration mechanics that shape every single request the model receives.
Here's the kicker: all of those components consume tokens before your user even types a word.
AI consultancy Systima ran a telling experiment: they took two popular harnesses — Anthropic's Claude Code and the open-source OpenCode — and ran identical tasks through both, using the exact same underlying model (Claude Sonnet 4.5). The results? Sharply different token overhead, driven entirely by differences in harness configuration. Same model. Same task. Wildly different economics.
This is not an edge case. A separate academic paper, "The Harness Effect: How Orchestration Design Sets the Token Economics of Enterprise Agentic AI," quantified the phenomenon at scale and arrived at numbers that should make any CIO sit up straight:
- 38% reduction in token consumption from harness optimization alone
- 41% reduction in cost per task
- 44% reduction in execution time
- All while maintaining comparable output quality
Let that sink in. Nearly half your execution time and more than a third of your costs could be a harness configuration problem — not a model problem, not a prompt problem, and not a budget problem. A configuration problem.
Nobody Is Watching the Meter
If the numbers above are alarming, the organizational reality behind them is almost comical. Ask your CIO whether your coding agent is rewriting its cache mid-session, and — as Broadcom SRE Advait Patel put it — "you will get a blank stare."
Here's why: agent configuration in most enterprises is distributed across teams that don't talk to each other.
"The harness is chosen by one team, the instruction file written by another, and the MCP servers attached by a third, so no single person sees the cumulative weight."
— Ashish Chaturvedi, Executive Research Leader, HFS Research
That fragmentation means there is no single owner of harness economics. The platform team picks the orchestration framework. The AI team writes the prompts. The infrastructure team plugs in MCP servers. Everyone is optimizing their own slice, and nobody is looking at what the agent is actually sending per request, how stable the cache prefix is, or what subagent fan-out costs at scale.
As HyperFRAME Research's Stephanie Walter noted, enterprises "need to measure the entire agent configuration, not assume model pricing tells them what an agent will cost." Right now, almost none of them do.
FinOps for Agents: Still in the Stone Age
Anyone who's been in enterprise IT long enough remembers the early days of cloud computing, when AWS bills would arrive and nobody quite understood what had generated them. It took years, and an entire discipline called cloud FinOps, to bring visibility and governance to cloud cost management.
Agentic AI is at the exact same inflection point — except it's moving faster and the costs are less transparent.
"FinOps for agents is where cloud FinOps was in 2013. Nobody has hired the equivalent of a cost optimization team focused on prompt engineering, harness configuration, and cache stability."
— Advait Patel, SRE at Broadcom
The discipline simply doesn't exist yet in most organizations. There are no standard metrics. There are no dashboards. There is no equivalent of the cloud cost anomaly alert that wakes someone up at 3 a.m. when an agent starts misbehaving and burning tokens at 10x the expected rate.
In production deployments — where agents run sustained workloads at scale rather than short demo sessions — that gap stops being an inconvenience and starts being a P&L problem.
The Scale Pressure Is Here, Right Now
This isn't a future problem. Gartner predicts that by 2028, 33% of enterprise software applications will include agentic AI — up from less than 1% in 2024. Meanwhile, 15% of day-to-day work decisions will be made autonomously by AI agents within the same timeframe.
Organizations aren't casually piloting agents anymore. They're deploying them into real workflows, with real workloads, where the cost dynamics are fundamentally different from a 10-minute demo. Proof-of-concept sessions are forgiving: they're short, usage is light, and accuracy is the only metric anyone evaluates. Production is unforgiving: workloads are sustained, scale is unpredictable, and token costs compound relentlessly.
That shift is forcing a new kind of question onto the IT agenda. Teams that were asking "which LLM should we use?" are now asking "why are our agents burning tokens we can't see?"
The Skills Gap Is Real, Too
The problem isn't just organizational — it's also a skills issue. The AI engineering community is starting to acknowledge that prompt engineering, while useful, is only the surface layer. The deeper craft involves understanding the full harness stack.
"As an AI Engineer. Please learn: Harness engineering, not just prompt engineering. Context engineering, not just long prompts. Prompt caching vs. semantic caching tradeoffs. KV cache management, eviction, reuse, and memory pressure at scale. Prefill vs. decode latency and why they optimize differently."
— @ParamSiddh (208 likes)
That sentiment is echoed widely in the practitioner community:
"Most people call themselves AI Engineers in 2026. Very few actually know what that means. Here is everything the real ones know that nobody teaches you. 01: Harness engineering, not just prompt engineering. Prompts are the surface. The harness is the system."
— @aisolram
The gap between "I can write a good system prompt" and "I can architect and optimize a production-grade agent harness" is enormous. And enterprises scaling agentic AI without closing that gap are essentially hiring racing drivers without teaching them how to tune the car.
What Good Harness Governance Actually Looks Like
So what does a mature approach to harness management look like? Analysts point to several emerging best practices:
1. Measure the whole configuration, not just model pricing.
Token cost per task must account for system prompts, tool schemas, MCP connections, subagent interactions, and cache behavior — not just the headline per-million-token rate from a model vendor.
2. Establish a single owner for harness economics.
Whether that's a new "AI FinOps" role or an expanded mandate for an existing platform team, someone needs to own the cumulative cost weight of the entire agent configuration.
3. Build runtime tracing and workload-level cost attribution.
Production deployments need the same observability infrastructure that mature cloud deployments have: execution analytics, cost baselines, anomaly detection, and forecasting. Without it, every scaling decision is a guess.
4. Treat harness configuration as a first-class engineering artifact.
Instruction files, tool definitions, MCP server selections, and caching strategies shouldn't be informal decisions made by whoever's available. They should be versioned, reviewed, and optimized with the same rigor as application code.
The Bottom Line
The model wars are largely over. Every major LLM at the enterprise tier is genuinely capable. The real competitive differentiation — and the real cost management lever — is now sitting one layer up, in the orchestration harness that most enterprises don't measure, don't own clearly, and don't optimize systematically.
The analogy to early cloud FinOps isn't just illustrative — it's a warning. Cloud cost chaos cost enterprises billions before the discipline matured. Agentic AI costs could follow the same trajectory, just faster and at greater scale.
The good news? The harness is configurable. The costs are reducible. The visibility gap is closable. But none of that happens without first acknowledging that the harness exists, matters enormously, and is currently being managed by nobody in particular.
That blank stare when you ask about cache rewriting? That's your starting point.
Published in Stream · Dispatch #456 · July 18, 2026 · 8 min read.
Reply to paolo@mont3.ch - every email gets a human answer within 24h.