Coding agents and legal agents work on documents with a very different composition. Coding functions can span multiple pages, have adverse effects, and complex dependencies. So can legal clauses and sections, but often in a very different way. For these tasks, we present some aggregate token consumptions, approximated by various usage bins.
Where agent spending goes #
| Source | Cursor | Lexifina |
|---|---|---|
| System & tool defs | 17% | 34.4% |
| Read tool outputs | 20% | 9.0% |
| Other tool outputs | 18% | 5.3% |
| Tool call arguments | 21% | 17.6% |
| Reasoning | 10% | 24.8% |
| Skills & plugins | 8% | 7.1% |
| Assistant text | 4% | 1.6% |
| User text | 3% | 0.37% |
Some insights into this utilisation #
System instructions explain how the agent should work. Tool definitions describe the actions it can take and the information each action needs. Lexifina spends more on this, presumably to provide more exhaustive user steering, and because our agents are primarily cloud-based.
Lexifina loads some tool definitions on demand. Definitions available to the model still occupy space in its request, alongside system instructions. Caching can reduce the cost of reading that text again, but it also puts pressure on the active context width. Rather, many secondary and tertiary tool calls are "delegated", advertised and discoverable to the agent but not actively loaded into context.
Tool outputs include the information returned by an agent after an action. Cursor assigns 20% of spending to read-tool outputs and 18% to other tool outputs. Lexifina's corresponding shares here are 9.0% and 5.3%. The primary user action involves redlining (marking up a paragraph) rather than raw document generation, and coding also generally consumes more context outside of complex legal work.
Skills and plugins account for 7.1% of spending here. A skill that is hard to discover can cause retries, so large definitions help the agent find the right action.
For Lexifina, useful efficiency means lower total cost across drafting, review, research, and matter summaries while preserving source coverage and reducing the corrections a lawyer must make. Time spent, failed actions, retries, and all delegated work belong in that assessment. A shorter request that leads to repeated repairs can cost more overall. In terms of raw token optimisation, we strive to make context reusable by keeping recurring instructions and tool descriptions in a consistent order, and by using cache boundaries to separate stable blocks from a growing conversation.