Agent reliability is the practice of keeping an AI agent correct in production after it ships: recording what it does, judging whether it did the job, finding the failures that repeat, fixing them, and making sure each fix holds. The terms below are the vocabulary of that practice, grouped in the order the work happens.
The harness
- Agent harness
- Everything around a language model that turns it into an agent: the system prompt, tools, memory, orchestration, workspace and checks. The model reasons; the harness decides what it sees, what it may do and whether the result counts. What is an agent harness?
- Harness engineering
- Designing and improving the harness rather than the model. Much of an agent's reliability comes from this work, because the same model behaves differently in different harnesses.
- Guides and sensors
- Two kinds of harness control. Guides steer the agent before it acts (instructions, tool descriptions, context). Sensors observe after it acts and help it correct (tests, validators, judges, review).
- Self-healing harness
- A harness with a closed loop from production back into it: failures in real traffic are found, named, judged, fixed and turned into regression tests, so a fixed failure cannot silently return. In Neens: the loop runs beside your harness, reading the traces it already emits.
- MCP (Model Context Protocol)
- An open protocol that lets an AI application call tools and read data from external servers. In Neens: the whole failure-to-fix loop is exposed as an MCP server, so a coding agent can read failures, pull a fix and report the pull request back. Evaluating agents over MCP
What gets recorded
- Trace
- One end-to-end run of an agent: a request comes in, the agent works, a response goes out. A trace is made of spans. In Neens: the raw unit you send, listed on the Traces page. Traces and sessions
- Span
- One step inside a trace: an LLM call, a tool call, a retrieval, a guardrail check or an agent step, with its inputs, outputs, timing and status. For model calls the span also carries the model name and token usage, which Neens turns into cost.
- Session
- A conversation-level rollup of traces. Traces that share a conversation id are grouped so a multi-turn chat reads as one unit.
- Tool call
- A tool or function the agent invoked, with its name, arguments, result, duration and any error.
- OpenTelemetry
- The open standard for emitting traces. In Neens: traces arrive over OTLP, OpenInference or raw JSON, and framework quickstarts show exporter configuration with no Neens SDK. Send traces
- Agent observability
- Recording and browsing what an agent did, step by step. It answers "what happened?" but not "was it right?" or "is it fixed?". Observability vs evaluation vs remediation
Evaluation
- Agent evaluation
- Deciding whether an agent's run was good, against criteria that reflect the job it exists to do. Evaluation turns a recording into a verdict.
- LLM-as-a-judge
- Using a language model to score another model's output against a rubric. In Neens: a judge is an automated evaluator, usually LLM-based and running on your own LLM connection, that can also be rule-based or composed into a composite. Judges have versions and deployments and run on demand, on a schedule or continuously. LLM-as-a-judge
- Score
- A judge's output for one trace or span: a metric, a numeric value, an optional pass or fail against a threshold, and a rationale explaining the verdict.
- Annotation
- A human label on a trace: a pass or fail verdict with an optional critique. Annotations are the ground truth judges are checked against.
- Judge alignment
- How often an automated judge agrees with human labels. In Neens: the Alignment view reports precision, recall, specificity and Cohen's κ per judge. Annotations and review
- Cohen's κ (kappa)
- A chance-corrected measure of agreement between two raters: 0 is roughly random, 1 is perfect. It is a stricter test of a judge than raw agreement, because two raters who both say "pass" to almost everything agree often by accident.
- Dataset and golden dataset
- A curated set of examples, built by hand, from a filter, from a cluster or from live traffic. A golden dataset is a frozen version used as the reference that gates and pre-production evaluations replay.
- Continuous evaluation
- Scoring a sample of live traffic as it arrives, so quality is measured all the time rather than in occasional test runs. In Neens: a primary score samples incoming traces, 5% by default.
Diagnosis
- Failure mode
- A named, defined kind of failure, the durable category that individual failing runs roll up into. Naming it is what turns a thousand bad conversations into one problem you can track.
- Failure clustering
- Grouping similar failing runs automatically. In Neens: failing traces are embedded, clustered and labeled with your LLM connection, giving a short, named list of what is going wrong. Failure clustering
- Issue
- A failure mode tracked as work, with a lifecycle from open through investigating and mitigating to resolved, the evidence sessions behind it, and the fixes it owns.
- Failure locus
- Where a failure actually lives: the agent's reasoning, a tool contract, a downstream or upstream service, or a control working as intended. The locus decides whether the right response is a change in your repository, a note to a service owner, or nothing at all.
Fixing and gating
- Remediation
- A typed, tracked fix for a failure, such as a prompt change or a tool fix, generated from trace evidence. In Neens: it moves from proposed to accepted, applied and then verified or regressed, and applying it requires a bound proof. Remediations
- Agent regression test
- A test built from a failure that already happened, run on every change so the failure cannot come back unnoticed. AI agent regression testing
- Regression set
- The growing collection of everything an agent must keep getting right. In Neens: every fix is verified against the failure's own evals and the accumulated regression set, and a human correction to a judge is captured into the set, so each correction makes the gate stronger.
- Eval gate
- An evaluation built from your own failures and used as a regression guard. In Neens: a gate ties a failure mode to a dataset of its real examples and a judge, and tracks the pass rate against a baseline. Eval gates
- Pre-production evaluation
- A whole-release check: replay a golden dataset against a candidate version, score it with the existing judges, and compare against a baseline before it ships.
- pass^k
- A rule that a candidate fix must pass k independent verification runs, all green, before it counts. In Neens: k defaults to 3, which defeats a fix that only passes sometimes. Eval-verified fix
- Human merge
- The rule that a person reviews and merges every fix. In Neens: the loop opens a pull request only when every run is green, and it never merges on its own at any autonomy level.
- Post-merge efficacy
- Measuring, after a fix ships, whether the failure actually dropped in production. In Neens: failure volume before and after each merged fix sets a verified or regressed verdict.
- Proving a cheaper model
- Running a cheaper candidate model through the same suite and the same judges as the current one, and switching only if it clears the same bar. Switching to a cheaper model without regressions