> ## Documentation Index > Fetch the complete documentation index at: https://docs.pandaprobe.com/llms.txt > Use this file to discover all available pages before exploring further. # Architecture > How PandaProbe captures, evaluates, and monitors agent behavior PandaProbe is a unified platform for the full agent development lifecycle. This page shows how its pieces fit together — from instrumenting your application to continuously monitoring agent quality in production. ## The agent engineering loop PandaProbe is built around a continuous loop: 1. **Trace** — capture what your agent actually did. 2. **Evaluate** — score those traces and sessions for quality and reliability. 3. **Monitor** — re-run evaluations on a schedule so regressions surface automatically. 4. **Heal** — feed those scores back to the agent itself, so it learns validated operating rules from its own failures. Each stage feeds the next: traces are the raw signal, evaluation turns that signal into scores, monitoring keeps those scores fresh as new data arrives — and the harness closes the loop by turning scores into behavior change, automatically. ## How data flows Add the PandaProbe SDK to your app. Wrap an LLM client, enable a framework integration, or annotate functions with decorators — see the [three layers of tracing](/tracing/overview). As your agent runs, the SDK emits **traces** (one logical run) made of **spans** (each LLM call, tool call, or step) to the PandaProbe API. The API queues incoming data and a background worker persists each trace and its spans, so ingestion stays fast and non-blocking for your application. Create an **eval run** to apply one or more **metrics** to selected traces or sessions. Workers compute scores asynchronously using LLM-as-judge metrics, embeddings, or deterministic aggregation. Save an eval configuration as a **monitor** to re-run it on a recurring cadence, so new traces and sessions are checked automatically. Optionally wrap the agent in the [self-healing harness](/harness/get-started/concepts): it evaluates every completed turn, posts a **diagnostic notice** when quality degrades, and the agent itself diagnoses the flagged traces and records a mitigation rule — which the harness validates before trusting. Inspect traces, drill into spans, and track scores and trends over time in the dashboard — or query everything through the API and CLI. ## Core building blocks A trace is one end-to-end run; spans are the tree of steps inside it. Group related traces under a session to see an agent's full lifecycle. Metrics score a trace or session; each result is stored as a score with a reason. Saved evaluations that run on a schedule to catch regressions in production. The harness posts diagnostic notices; the agent answers with validated operating rules. The operational envelope that turns evaluation scores into agent behavior change. ## Two ways to connect PandaProbe separates how applications send data from how people manage their workspace: | Plane | Who uses it | How it authenticates | | -------------------- | -------------------------------------------------------------- | ----------------------------------- | | **Data plane** | SDK clients and the CLI sending traces, spans, and evaluations | Org-scoped API key + project name | | **Management plane** | The dashboard, for users, projects, and billing | Sign-in with your identity provider | The project an API key writes to is resolved by name within your organization, so SDK clients only need an API key and a project name to start sending data. ## Deployment options Managed deployment with a free tier — no infrastructure to run. Sign up and start tracing. Run the open-source stack yourself with Docker Compose and keep your data in your own environment. ## Next steps Trace your first LLM call in under 2 minutes. Explore the three layers of instrumentation. Learn how metrics, scores, and monitors work. Turn a traced agent into a self-healing agent.