Home / Concepts /Model Spends, Harness Budgets

Model Spends, Harness Budgets

Agent architecture is the design of bounded resource allocation under uncertainty. The model spends. The harness budgets.

Definition

The model is an open-ended spender — of context window, tokens, latency, side effects, money. The harness is the engineered ceiling that decides what the model is allowed to spend, on what, when, and under whose authority. The interesting design work happens in the harness, not in the prompt.

The problem it names

Two years of agent incidents in production have a recurring shape: almost none of the failures are prompt-writing failures. They are harness-missing failures. Engineers spend the bulk of their attention on prompts because prompts are visible in the IDE — readable, editable, with feedback in seconds. When an incident actually fires, the postmortems read like the same sentence over and over:

• A customer question triggers a retry loop that burns a week of token budget in a single conversation. No upper bound.
• An agent granted bash for "code analysis" eventually runs rm -rf on a development tree. No tool scope.
• A data-repair agent with production write access corrupts three tables in thirty minutes. No human-in-the-loop on destructive operations.
• A long-running agent's context grows to 180k tokens and its reasoning begins to drift. No auto-compaction or reset policy.
• An agent hallucinates a tool result; every downstream action runs on the fabricated data; the incident surfaces six hours later. No side-channel verification.

Lined up, the pattern is unambiguous: the parts that got iterated never failed; the parts that were never designed broke everything. Prompts got tuned, so prompt-layer failures were absorbed early. The harness boundaries — budgets, scopes, blast radius, escalation, observability — were not in the IDE and so were never explicitly designed. The incidents leaked through there.

This imbalance has a precedent. Early web engineering spent its attention on CSS and JS while outages came from missing observability. DBAs tuned queries while production incidents came from absent connection-pool limits. The visible work gets polished. The infrastructure work gets neglected, until it wakes someone up at 3 AM. Agent systems are at the same stage now. This claim names that imbalance.

How it changes design decisions

Once you accept that the model is a spender, every architectural decision becomes a budget decision. Context window is not "how much can the model read" but "how do I allocate scarce token capacity across competing inputs." Tool scope is not "what can the agent do" but "what surface am I willing to let it touch." Reasoning depth is not a quality knob but a cost knob. The job of the harness is to set those ceilings before the model gets to decide.

Lineage

Not new. The pattern recurs across decades of systems engineering:

• OS process scheduling — processes want all the CPU; the scheduler decides who gets it
• JVM garbage collection — application allocates freely; GC budgets pause time and heap
• CDN rate limiting — clients want unlimited bandwidth; the edge sets per-tenant ceilings
• Linux cgroups — workloads share a host; the kernel enforces CPU / memory / IO budgets
• SRE error budgets — teams want to ship freely; the budget gates release velocity

The agent-era version says: the LLM is the workload; the harness is the kernel.

What it is NOT

• Not a critique of LLMs. The model being a spender is a fact of its design, not a failure.
• Not "the harness replaces the model." The harness is uninteresting without something to budget for; the model is dangerous without something to budget it.
• Not just observability. Observability tells you what was spent. Budgeting decides what is allowed to be spent.
• Not prompt engineering. The prompt sits inside one cell of the harness; the harness designs the cell.

Where it's developed further

• Position paper: A Two-Dimensional Framework for AI Agent Design Patterns (Huang & Zhou, 2026)
• Book treatment: Designing AI Agents (Manning, 2026) — runs as the through-line of the whole book; explicit in Ch1, Ch7 governance layer, Ch10 production harness
• Concrete patterns: G2 Blast Radius Control, G5 Observability Harness, P1 Context Triage

Cite as: Huang, J. (2026). Model Spends, Harness Budgets. kage-ai.com/concepts/model-spends-harness-budgets/