Agent benchmark

A benchmark designed for systems that plan, call tools, and act across multiple steps. It usually scores task completion, tool use, state changes, recovery behavior, and trace quality rather than one final answer.

Agent harness

The runtime scaffold around an agent: prompts, tools, memory, policies, retry logic, and execution loop. A harness can strongly affect eval results, so model comparisons should control or report it.

Agent skill

A folder-packaged unit of agent capability containing instructions, scripts, and reference material, loaded on demand rather than held in context. Progressive disclosure keeps only the metadata resident until the skill triggers, which is what lets one agent carry many specialisms.

Agentic RL

Reinforcement learning applied to agents acting over long horizons with tools and environments, framed as a partially observable problem rather than a single-step one. Credit assignment, reward design, and rollout infrastructure all become harder than in single-turn RLHF.

Alignment

The process of making a model or agent behave according to intended goals, human values, and safety constraints. In evals, alignment is tested through edge cases, adversarial prompts, policy scenarios, and real user workflows.

Calibration

How well a system's confidence matches actual correctness. A calibrated model is uncertain when it is likely wrong, which matters for escalation, human review, and high-stakes decisions.

Canary task

A small, stable eval task used as an early warning signal. Canary tasks should be cheap, frequently run, and sensitive to regressions in important behaviors.

Compaction

Summarizing or rewriting accumulated context mid-run so a long-horizon agent can continue past its context window. Done badly it silently discards the detail a later step needed, which is why compaction quality is itself worth evaluating.

Context engineering

The practice of curating the minimal high-signal token set at each inference step: just-in-time retrieval, compaction, tool design, and sub-agent isolation. It is the successor framing to prompt engineering once the input is assembled dynamically rather than written once.

Context rot

The measured degradation in accuracy as input length grows, appearing well before the context window is full and worsening with distractors or low query-to-target similarity. It is the empirical reason to budget context deliberately instead of filling the window.

Data flywheel

A loop that converts production traces and human corrections into structured evals, and those evals into targeted fixes. The defining property is that operating the system generates the signal that improves it, rather than improvement depending on new model releases.

Deterministic grader

A script, rule, unit test, database query, or exact verifier that scores outputs or final state without model judgment. It is preferred when success can be objectively checked.

Eval harness

The infrastructure that runs tasks, resets environments, captures traces, calls graders, and aggregates results. A good harness makes evals repeatable, auditable, and comparable across model or prompt changes.

Failure taxonomy

A structured list of ways a system fails, such as wrong reasoning, bad retrieval, unsafe tool use, formatting errors, or missed constraints. Taxonomies turn vague failures into fixable categories.

Generative reward model

A model that produces a reasoned judgment used as the reward signal for training another model, instead of a scalar learned from preference pairs. It scales to domains with no programmatic verifier, at the cost of inheriting judge bias into the training loop.

Golden set

A curated set of examples with trusted labels, expected answers, or target states. Golden sets are often used for regression testing, judge calibration, and manual review of high-value behaviors.

Ground truth

The reference answer, label, target state, or verified external outcome used to decide whether a system succeeded. For agents, ground truth may be a database state, file diff, or transaction result.

GRPO

Group Relative Policy Optimization: a policy-gradient method that compares a group of sampled trajectories against each other rather than against a learned value baseline. Because it needs only relative ranking within a group, it pairs naturally with LLM judges as reward functions.

Hallucination

A fluent but unsupported, fabricated, or false output. In agent systems, hallucination can also mean claiming that a tool action succeeded when the external state did not actually change.

Harness search

Automated outer-loop optimization over the harness code itself — tools, prompts, control flow, orchestration — evaluated end to end. It treats the scaffold as the thing being learned, in contrast to hand-tuning a fixed scaffold around a fixed model.

Inference-time scaling

Spending more compute at inference — sampling, verification, iterative refinement — to raise accuracy without retraining. Adaptive variants allocate that budget per problem, spending heavily on hard cases and exiting early on easy ones.

Inter-rater agreement

The degree to which multiple human graders give the same labels or scores. Low agreement often means the rubric is ambiguous, the task is subjective, or more grader training is needed.

Judge model

An LLM used to grade another model or agent. Judge models scale open-ended evaluation, but they need calibration against humans or deterministic checks because they can be biased or inconsistent.

Model Context Protocol (MCP)

An open protocol for exposing tools, data, and prompts to an agent through a uniform server interface. Its practical significance is standardization: capabilities become portable across hosts instead of being reimplemented per integration.

Pass@k

The probability that at least one of k attempts succeeds. It is useful when retries, sampling, or best-of-k selection are allowed, and it measures potential capability more than consistency.

Pass^k

The probability that all k attempts succeed. It is stricter than pass@k and better reflects reliability for customer-facing workflows where repeated attempts should all work.

Progressive disclosure

Loading tool definitions, skills, or documentation only when they are needed rather than holding all of them in context. It is the main defense against tool-count growth eating the context budget and degrading tool selection.

Prompt injection

A security and reliability failure where untrusted text attempts to override instructions, leak secrets, manipulate tools, or change the agent's goal. It is especially important in browser and retrieval systems.

Recursive self-improvement

A system that modifies its own code, prompts, or scaffold and validates each change empirically, so improvements compound across iterations. The binding constraint is the evaluator: gains cannot exceed the reliability of the signal used to accept changes.

Red team eval

An evaluation that actively searches for harmful, unsafe, or policy-violating behavior. Red teaming often uses adversarial prompts, malicious documents, unusual workflows, or expert testers.

Reflective prompt evolution

Optimizing prompts by having a model read execution traces, diagnose failures in natural language, and propose revisions, rather than following a numeric gradient. Traces carry far more signal than a scalar reward, which is why it can be dramatically more sample-efficient.

Reliability

How consistently a system succeeds across repeated trials, varied inputs, and realistic operating conditions. Reliability is different from peak capability: a model can sometimes solve hard tasks yet fail often.

Retrieval eval

An evaluation of whether a system finds, ranks, and uses the right source documents. It can measure recall, precision, citation faithfulness, and whether retrieved context actually improves answers.

RLVR

Reinforcement learning from verifiable rewards, where correctness is checked by a program, test, or proof rather than a learned preference model. It resists reward hacking better than learned rewards, but only covers tasks whose success can actually be verified.

Rubric

A written grading standard that defines success, partial credit, failure, and edge cases. Good rubrics reduce grader ambiguity and make model-based or human judgments easier to audit.

Safety eval

An evaluation focused on harmful behavior, policy violations, misuse risk, privacy leakage, prompt injection, unsafe tool use, or failure to refuse dangerous requests.

Tool

An external capability an AI system can call, such as search, code execution, browser control, file access, APIs, calendars, databases, or messaging systems.

Trace

The recorded sequence of messages, tool calls, observations, errors, decisions, and state changes during a run. Traces help explain why an aggregate score changed or why an agent failed.

Trajectory

The full path an agent takes through a task, including planning, intermediate steps, tool use, recoveries, and final outcome. Two agents may reach the same answer through very different trajectories.

Worst-case eval

An evaluation designed to expose rare but severe failures rather than average behavior. It is useful for safety, security, and reliability work where tail risk matters.