Traffic
Messages, model, app identity, tokens, latency, cost, and completion.
// AI agent observability
Punk records model traffic, tools, cost, latency, policy decisions, and selected routes as one replayable history. The same evidence can support debugging, evaluation, and carefully gated adaptation.
// one execution story
Messages, model, app identity, tokens, latency, cost, and completion.
Declared actions, inputs, outputs, side-effect posture, and approval state.
The selected path, fallback behavior, and a reason attached to every response.
Feedback and application signals when your system supplies them.
// from traces to adaptation
Punk is not positioned as a replacement for every evaluation or monitoring product. It adds an execution layer that can use observed patterns to propose, test, and route through reusable paths.
The configured provider remains the source of truth while Punk establishes the workload baseline.
Find request families, stable tool plans, candidate caches, and patterns that should remain live.
Replay and shadow comparison reveal mismatches within a defined scope; they do not prove universal correctness.
Eligible requests can use a verified path. Uncertain or ineligible work returns to the live provider.
// choose the right stack
Trace exploration, debugging, evaluation, quality review, and operational monitoring.
Provider access, routing, credentials, quotas, retries, and other traffic controls.
Evidence-gated reuse, policy-aware execution, live fallback, and routes that improve from production experience.
Deployment varies by stack. Punk can receive OpenAI-compatible model traffic with an endpoint change. Agent tools, identity, outcomes, and richer governance may require additional integration.
Observe first. Decide what deserves deeper evaluation only after the traces are complete enough to support it.