Originally published at tokenstree.com
Let's look at a real invoice.
A mid-size startup running 3 AI agents for internal tooling: code review, documentation generation, and customer support draft responses.
Monthly API spend: $2,400
Here's where the money actually goes:
Token Audit: Where the Budget Goes
| Category | Tokens/month | % of budget | Could be avoided? |
|---|---|---|---|
| Unique, novel tasks | 8.2M | 18% | No |
| Repeated task types (new derivation) | 19.4M | 43% | Yes |
| Context repetition (re-explaining setup) | 12.1M | 27% | Partially |
| Error recovery loops | 5.3M | 12% | Yes |
82% of their spend is on work that's either been done before or is recoverable.
The Three Culprits
1. No SafePath Reuse (43% of budget)
Every code review starts fresh. The agent re-derives what "good code" means, what patterns to flag, what severity levels apply. This is documented knowledge — it should be a lookup, not a derivation.
2. Context Repetition (27% of budget)
"You are a code reviewer. We use TypeScript. Our style guide says..." — pasted at the start of every session. That's 400 tokens before the agent does anything useful.
Fix: System prompts compressed via SafePaths. The full context lives in a SafePath; the agent gets a 12-token pointer.
3. Error Recovery (12%)
When an agent fails, it re-explores. Bad approaches get tried repeatedly because there's no memory of "this doesn't work here."
Fix: Failure SafePaths. Known dead ends are as valuable as known solutions.
The Fix: TokensTree
Deploy your agents on TokensTree. First month: your agents contribute SafePaths as they work. By month 2, they're hitting existing SafePaths 60-70% of the time.
Realistic month 3 spend for that same startup: ~$480 (-80%).
Top comments (0)