Retries multiply
An agent can turn one failed task into many billable model calls before a person sees the failure.
Agents retry, loop, and run in the background. CachePilot gives the humans responsible for the bill a governed proxy with budgets, policy, receipts, telemetry, and attribution.
The buyer fear is concrete: invisible spend, no attribution, and no receipt trail when an agent makes hundreds of decisions on its own.
An agent can turn one failed task into many billable model calls before a person sees the failure.
Scheduled and long-running automation keeps spending when nobody is watching a terminal.
Shared provider keys make it hard to attribute spend to a project, route, policy, or agent workload.
Keep the agent workflow and provider key. Route calls through CachePilot to establish a control and evidence layer around every request.
Apply project policy to cap generated tokens instead of relying on every caller to enforce a limit.
Observe first, then govern request shape, tools, models, and expensive behavior at the proxy.
Return X-CP-* proof with each response and record content-free telemetry by project, model, and route.
Separate cached and uncached tokens so operators can see whether repeated agent context is saving money.
/v1/messages. An OpenAI-compatible Claude/Fable adapter receives full governance when it targets CachePilot's governed Responses surface; the Chat Completions compatibility route is telemetry-only today.Receipt headers let the calling agent, application, logs, and operators inspect what governance was applied without storing prompt or output content by default.
< HTTP/2 200 < X-CP-Receipt-Id: b7a3c8e2-1f4d-4a9b-8c6e-3d5f7a2b1e09 < X-CP-Policy-Version: 3 < X-CP-Output-Budget-Applied: 4096 < X-CP-Skills-Applied-Hash: a3f829c1e2b4d6f8 < X-CP-Prefix-Hash: 7b2df491a8c3e5d7 < X-CP-Mutations: output_budget,cache_breakpoint