A chatbot costs what a user's conversation costs. An agent running unattended costs whatever you allow it to cost, and the difference between a well-budgeted agent and a poorly budgeted one is routinely a factor of three or four on the monthly bill. Here are four patterns Australian teams running Claude agents in production use to keep that number predictable.
1. Scope context per action, not per session
The default pattern for an agent is to accumulate context as it works, so by step twelve it is carrying everything from steps one through eleven whether or not any of it is relevant. A pre-tool-use hook that filters context down to what the current action actually requires is the single highest-value change available, and teams applying it consistently report cutting per-request token use by more than half.
The counterintuitive part is that this usually improves output quality as well as cost. A model asked to act on 11,000 relevant tokens makes better decisions than one asked to act on 37,000 tokens of which most are noise from earlier steps.
2. Match model tier to task tier within a single workflow
Teams tend to pick one model for an entire agent workflow, which means the most expensive model in your lineup handles the trivial classification steps as well as the genuinely hard reasoning. Routing within a workflow, using a lighter model for extraction and classification and reserving the stronger model for the steps where judgement matters, is straightforward to implement and typically cuts spend substantially.
Classification, extraction and routing decisions: lighter model, low effort setting.
Drafting and summarisation of routine content: middle tier.
Judgement calls, ambiguous cases, anything customer-facing or financially consequential: strongest model available.
3. Use prompt caching for the parts that never change
Most agent workflows carry a large stable prefix: system instructions, tool definitions, business rules, reference material. Cached input tokens cost a small fraction of the base rate on repeat reads, so any workflow that sends the same substantial prefix hundreds of times a day is leaving real money on the table by not caching it. For a high-volume workflow this alone can cut input costs dramatically.
The discipline required is structural: put everything stable at the front of the prompt and everything variable at the end. Teams that interleave stable and variable content get much less benefit, and this is worth checking in an existing implementation because the ordering is easy to get wrong without noticing.
4. Set hard budget ceilings, not just monitoring
Monitoring tells you about an overspend after it happens. For an agent that runs autonomously, particularly one triggered on a schedule or by inbound volume, a hard ceiling that halts execution and alerts a human is the control that actually prevents a bad month. A Brisbane team we advised discovered a retry loop that had quietly consumed roughly $4,700 AUD over a long weekend before anyone looked at the dashboard on Tuesday.
Set ceilings per workflow rather than per account, so one runaway process cannot consume the budget the rest of the business depends on, and make the alert go to a human who is actually on call rather than a shared inbox nobody reads on a Saturday.
Measuring before optimising
All four patterns assume you know where your tokens are going, and most teams do not at the point they first get a surprising invoice. Before changing anything, instrument the workflow so you can attribute spend by step: which action in the agent loop consumes the most tokens, how often each step runs, and what proportion of context is stable versus variable. That instrumentation usually takes a day and frequently makes the priority obvious, because one step is responsible for most of the cost.
It also protects against optimising the wrong thing. We have seen teams invest a fortnight in prompt caching for a workflow whose real problem was an unbounded retry, and teams rewrite their routing logic when a single oversized context payload accounted for most of the bill. Measure first, then pick the pattern that addresses what you actually found.
What this adds up to
Applied together on a workload spending around $12,000 AUD a month, these four patterns routinely bring the bill down to somewhere between a third and a half of that without reducing what the agent actually does. The savings come from removing waste, not capability, which is why they rarely involve a quality trade-off worth arguing about.
The broader point for Australian teams putting agents into production is that inference cost is an engineering property, not a vendor property. Teams who treat their bill as something the vendor sets tend to have unpredictable costs. Teams who treat it as something their architecture determines tend to have predictable ones, and the difference between those two positions is about a week of deliberate work.
