Cmd+K, then type “glossary”).
A2A
Agent-to-agent: one agent calls another over a shared protocol. A2A (agent-to-agent) is the protocol agents use to hand tasks to other agents. Zespan records each handoff as a span, so you can see which agent asked which, what it sent, and how the task ended. See Agents.Agent
An AI program that plans, calls tools and takes several steps to finish a job. An agent is a run of your AI app that decides what to do next: it calls models, uses tools and may hand work to other agents. Zespan groups its traces by agent name so you can watch each one’s cost, errors and flow. See Agents.Alert rule
A threshold on cost, errors or scores that notifies your channels when it is crossed. An alert rule checks a metric every few minutes (cost, error rate, latency, an eval score) and notifies the channels you chose when it crosses your threshold. It can be scoped to one environment. See Alerts.Annotation queue
A list of traces for people to review and label by hand. An annotation queue collects traces for your team to review. Reviewers add labels and notes, and those labels can calibrate evaluators or become dataset rows. See Annotation queues.Change
A recorded edit to your AI system, such as a prompt deploy, policy edit or model swap. A change is something that altered how your AI behaves: a prompt version going live, a policy edit, a model swap or a deploy you report from your pipeline. Zespan lines changes up against metrics so you can see what a change did. See Changes.Cost attribution
Spend split by team, feature, user or any tag you send. Cost attribution breaks your AI spend down by the tags and ids you attach to traces, such as team, feature or end user, so you can see who and what the money went to. See Cost attribution.Dataset
A saved set of inputs, with expected outputs if you have them, used to test your AI. A dataset is a collection of inputs (and optionally expected outputs) that you run evaluators or simulations against. Build one by hand, upload it, or pull rows from real traces. See Datasets.End user
A person using your AI app, identified by the user id you pass to the SDK. An end user is the customer of your own app, not a Zespan teammate. Pass a user id in the SDK context and Zespan shows traces, cost and errors per user. See Users.Environment
A label such as production or staging that separates traces from different deployments. An environment tags traces with where they ran, for example production or staging. Filters and alerts can then be scoped to one environment. See Settings → Environments.Evaluation run
One batch of scores from running evaluators over a dataset or a sample of traces. An evaluation run is a single pass of one or more evaluators over a dataset or a sample of traces. It ends with a set of scores you can compare to earlier runs. See Evaluations → Runs.Evaluation template
A reusable recipe for an evaluator; deploy it to create an evaluator on this project. An evaluation template is a ready-made definition of a check, such as relevance, toxicity or correctness. Deploying it creates an evaluator on your project that you can then tune. See Evaluations.Evaluator
A check that scores traces (an LLM judge or a rule) and runs on new traces automatically. An evaluator scores your AI’s outputs, either with an LLM judge or with a rule. Once deployed it scores new traces on its own, at the sample rate you set, and its scores show up on traces, in Issues and in alerts. See Evaluations.Guardrail
A check that runs on a request or reply in real time and can block, redact or warn. A guardrail runs while a request or reply is in flight. Depending on how you set it, it can block the message, redact part of it or only warn. Every hit is recorded so you can review it. See Guardrails.Incident
A live problem that needs a response now, opened when Zespan detects an anomaly or an issue keeps recurring. An incident is opened when Zespan sees a sudden spike or drop against the usual baseline, or when an issue keeps coming back. It tracks status, timeline and what changed around it, and links to the issues behind it. Write a postmortem once it is resolved. See Incidents.Issue
A recurring problem Zespan found in your traces, grouped by cause, with a fix and a way to prevent it. An issue is a pattern of failures or low scores that Zespan grouped from many traces. It states the likely cause with the evidence behind it, a suggested fix and a way to stop it coming back. See Issues.Layer
A part of your stack a trace passes through, such as the model, a tool or retrieval. A layer is one part of your AI stack, for example the model, a tool, retrieval or a gateway. Zespan tags spans by layer so it can show where in the stack something broke. See Traces.Lens
A way of reading one trace: for example flow, cost, or time. A lens is a view of a single trace that answers one question, such as where the time went, where the money went, or how the steps connect. Switch lenses without leaving the trace. See Traces.LLM connection
A provider key you save so Zespan can call a model for judges and tests. An LLM connection stores your own provider key (OpenAI, Anthropic and others) so Zespan can run LLM judges, simulations and the playground with your account.MCP
Model Context Protocol: a standard way for AI apps to use tools and data sources. MCP (Model Context Protocol) lets an AI app discover and call tools on a server. Zespan records MCP calls as spans, and offers MCP keys so assistants such as Claude or Cursor can read your project. See Tools.Policy
A written rule set that Zespan turns into guardrails and tests, with versions and a review step. A policy is a plain-language set of rules for how your AI should behave. Zespan turns it into guardrails and test cases, keeps versions, and asks for a review before anything goes live. See Policies.Prompt label
A movable name like production that points to one version of a prompt. A prompt label is a pointer such as production or staging. Your app asks for a prompt by label, and moving the label to another version changes what it gets without a code deploy. See Prompts.Prompt version
One saved revision of a prompt; every edit makes a new version. A prompt version is a fixed snapshot of a prompt’s text and settings. Versions never change once saved, so you can compare them and roll back safely. See Prompts.Retroactive eval
Scoring traces you already have, instead of waiting for new ones. A retroactive eval runs an evaluator over traces that were recorded before it existed, so a new check has history to compare against straight away. See Evaluations.Root cause
The most likely reason an issue happens, with the evidence that points to it. A root cause is Zespan’s best explanation for why a group of traces failed. It cites the traces, spans and changes it based that on, and says how sure it is. See Issues.Sampling
Keeping only some traces to save cost; the rest are dropped. Sampling keeps a share of your traces and drops the rest, at the sampling rate you set when you start the SDK. It lowers volume and cost, but sampled-out traces cannot show up in issues or scores.Session
A group of traces that belong to one conversation or visit, tied by a session id. A session is a run of related traces, such as one chat conversation, tied together by the session id you pass to the SDK. It lets you read a whole conversation instead of single calls. See Sessions.Simulation
A replay of scenarios against a new model or prompt to compare results before you ship. A simulation replays real or written scenarios against a different model, prompt version or agent, then compares the outputs so you can catch regressions before they reach users. See Simulations.Span
One step inside a trace, such as a model call, a tool call or a retrieval. A span is a single timed step in a trace: a model call, a tool call, a retrieval or a guardrail check. Spans nest, so a trace reads as a tree of what your app did. See Traces.Tool call
When an AI model asks your code to run a function, such as a search or a database lookup. A tool call is a request from a model to run one of your functions. Zespan records the name, arguments, result and duration, and flags the ones that fail. See Tools.Trace
The full record of one request or agent run, from the first input to the final output. A trace is everything that happened for one request or run: every model call, tool call and step, with timing, cost and outcome. Open one to see exactly what your AI did. See Traces.Trajectory
The path an agent took: the ordered steps and tools it used to reach an outcome. A trajectory is the sequence of steps an agent took in one run. The Trajectory lens on a trace grades that sequence, so a run can fail on its steps even when its final answer passes. See Traces.Verdict
The outcome Zespan gives a run: did it do what it should. A verdict is the result of a check on a run, such as pass or fail, along with the reason. Simulations show one per scenario, and Issues are grouped from traces whose verdict is failed or degraded. See Simulations.Next steps
Core concepts
How events, spans and traces fit together.
Dashboard overview
A tour of the pages these words come from.