Gateways meter AI traffic. Gesta observes agent behavior.
Your gateway answers one question well: who called which model, and when. The questions that decide budgets, audits, and incident reviews — what did the agent do, what did it produce, what did it actually cost — only exist on the endpoint, where the agent runs.
Two line items can't price an agent.
Gateway cost reports multiply two numbers: input tokens and output tokens. Agent sessions replay the same context on every turn — so the tokens that dominate the bill are the cache tiers the gateway's arithmetic never sees.
POST /v1/messages 200 model frontier-large input_tokens 1,842 output_tokens 697 — — — — — — — — — — — — — — cost = in × rate + out × rate = $0.016
Agent workloads are cache-dominated. Price only input and output, and the bill is wrong in its first digit. Gesta prices every token tier.
Data shown is illustrative
Spend is not output.
A proxy sees traffic — requests, bytes, latency. And model output tokens aren't output either: most of a session is the agent thinking, reading, retrying. Nothing in a request log connects a dollar of spend to work actually produced.
14:02:11 POST /v1/messages 200 2.1 MB 1,204 ms 14:02:19 POST /v1/messages 200 3.4 MB 980 ms 14:02:31 POST /v1/messages 200 1.7 MB 644 ms 14:02:48 POST /v1/messages 200 2.9 MB 1,411 ms — — — — — — — — — — — — — — requests 2,417 · bytes out 1.9 GB
- Code authored412 lines
- Tests authored96 lines
- Docs authored1,024 words
- External tools17 MCP calls
- Session cost$12.40
A proxy knows what you spent. It can't know what you got. Gesta measures what agents actually author — and a local agent writing files never crosses a proxy at all.
Data shown is illustrative
A request log is not an audit trail.
When the audit comes, “someone called a model at 09:41” is not an answer. And an agent that talks to a local tool or calls an API directly never traverses your proxy at all — the gateway's log has no row to show.
09:41:07 key prod-07 POST /v1/messages 200 proves: someone, holding this key, called a model at 09:41. cannot answer: • is this line of code AI-written? • what left in the prompt body? • what about calls that never routed through the gateway?
Request logs prove someone called a model. They can't tell you whether a line of code is AI-written, or whether customer data left in a prompt that never touched your gateway. Those answers only exist where the agent runs. That's where Gesta sits.
Data shown is illustrative