Alexey Shurov.
Glossary

Production AI glossary, the terms that matter once an agent is live

Plain definitions for the words people use about AI agents, automation and bottlenecks, written from building and running these systems rather than from a textbook.

Updated 25 September 2026 . 44 terms . Alexey Shurov

Agentic AIAI agentBatch APIBottleneckCatalogue mappingContext windowCycle timeDead letter queueDORA metricsDriftEmbeddingsError budgetEvalEvent logFlow efficiencyGroundingGuardrailsHallucinationHuman in the loopIdempotencyIntelligent document processingLittle's LawLLM as a judgeMulti-agent systemObservabilityOCROrchestratorPickup timeProcess miningPrompt cachingQueue timeRAGReplay regressionRetry with backoffRollbackRules engineShadow modeStructured outputThroughputTokenTool callingTouch timeTracingWork in progress

Agentic AI

A way of building AI systems where models plan and act across several steps and tools instead of producing a single reply. The useful question is not whether a system is agentic but which decisions it makes on its own, which it recommends, and which stay with a person.

AI agent

Software that uses a language model to decide what to do next, calls tools such as a database, an email inbox or an ERP, and works through a task in several steps rather than answering one prompt. In production an agent is only as good as the checks around it, what it is allowed to touch, and who reviews what it cannot settle.

Batch API

A way to send many requests at once for processing within hours rather than seconds, usually at a lower price. It suits overnight document runs and backfills, not anything a person is waiting on.

Bottleneck

The step in a process with the least capacity or the longest wait, which sets the pace for everything after it. Fix it and the bottleneck moves downstream, so the work is to find the next one before it bites.

Catalogue mapping

Matching the way customers or suppliers describe an item, with their own names, part numbers, units and pack sizes, to your own item codes. In order automation it decides accuracy more than the reading does.

Context window

The maximum amount of text, measured in tokens, that a model can take in for one request, including instructions, documents and conversation. A larger window costs input tokens on every call, so sending only what the task needs keeps cost and errors down.

Cycle time

The time from when work on an item starts to when it is finished. Breaking it into stages shows where the time goes, which is usually waiting rather than working.

Dead letter queue

A holding place for work items that failed repeatedly, so they can be inspected and replayed instead of being lost or retried forever. It turns silent failures into a visible list someone owns.

DORA metrics

Four software delivery measures from the DevOps Research and Assessment programme, deployment frequency, lead time for changes, change failure rate and time to restore service. Useful as averages, but they do not show which stage the work waits in.

Drift

When an agent's quality changes over time without anyone changing it, because the model was updated, the inputs changed or new edge cases appeared. A weekly review of a sample of real outputs is the simplest way to catch it.

Embeddings

Numeric representations of text that place similar meanings close together, used to search by meaning rather than exact words. They power most retrieval and matching, such as finding the catalogue item closest to a customer's description.

Error budget

An agreed number of wrong outputs a system may produce in a period before it is paused or fixed. It turns a vague expectation of accuracy into a rule everyone knows in advance.

Eval

Short for evaluation. A repeatable test of how well a model or agent does a task, run against a fixed set of cases with known good answers. Evals are what let a team change a prompt, a model or a rule and know whether things got better or worse.

Event log

A record of what happened to each case and when, such as order created, approved and shipped. Process mining, cycle-time dashboards and most bottleneck analysis depend on having one.

Flow efficiency

Touch time divided by total cycle time. A low figure means most of an item's life is spent waiting, so the fastest improvement is removing waits rather than working faster.

Grounding

Tying every claim an AI makes to a source it was given, such as a document, a database row or a web page, so a reviewer can check it. Grounded output that cites its sources is far easier to trust and to audit than fluent output that does not.

Guardrails

Checks placed around a model's input and output to keep it within bounds, such as refusing certain requests, validating formats or blocking actions above a limit. Guardrails reduce risk but do not replace testing the agent on real cases.

Hallucination

When a model states something that is not supported by its input or by fact, often confidently. No model is free of it, so production systems reduce the chance with grounding and validation, and catch what gets through with checks and review.

Human in the loop

A design where a person reviews or approves an agent's output before it takes effect, usually at the decision that carries the risk. Done well it is a deliberate handoff with the agent's reasoning and sources attached, not a person rubber-stamping every step.

Idempotency

The property that running the same operation twice has the same effect as running it once. For agents that write to real systems it is essential, because retries happen, and a retry must not create a second order or pay someone twice.

Intelligent document processing

Software that reads documents such as invoices, purchase orders and timesheets and turns them into structured data. Reading is the easier part. Validating totals, matching lines and routing exceptions decide whether the output can be trusted.

Little's Law

The rule that average time through a process equals work in progress divided by throughput. It explains why carrying less work in progress makes each item finish sooner without anyone working faster.

LLM as a judge

Using a language model to grade another model's output against a rubric, often used when answers are too varied to compare word for word. It scales well but needs its own checks, since the judge can be wrong in consistent ways.

Multi-agent system

Several agents, each with one job, passing work between them, for example a reader, a checker, a reconciler and a reporter. Splitting the work makes each step testable and makes failures visible, at the cost of more moving parts. One well-scoped agent is often the better first version.

Observability

Being able to see what an agent did and why, from its inputs, tool calls, sources and outputs to cost and latency, for every run. Without it, a failure next month is a guess. With it, it is a lookup.

OCR

Optical character recognition, turning an image of text into text. Modern systems often combine OCR with language models, but even very accurate character recognition can produce wrong line items without validation on top.

Orchestrator

The part of a multi-agent system that routes work between agents, tracks state and decides what happens after each step. It is usually deterministic code, not a model, because routing and retries need to behave the same way every time.

Pickup time

In software delivery, the time from a pull request being opened to its first review. It is often the single largest wait in a delivery pipeline.

Process mining

Reconstructing how a process actually runs from the event logs in systems such as an ERP or CRM, to see where cases wait, loop back or get abandoned. It needs clean event data, which many mid-size companies do not have yet.

Prompt caching

A discount providers give when the start of a prompt repeats across calls, such as long instructions or a reference document. It can cut input costs sharply for agents that send the same context many times.

Queue time

The time an item spends waiting between steps, for a person, an approval or a reply. It is usually the largest part of cycle time and the cheapest part to shorten.

RAG

Retrieval augmented generation. The system first searches your own documents for relevant passages, then gives them to the model to answer from. It keeps answers current and citable without retraining a model, and its quality depends mostly on the search step.

Replay regression

Testing a new version of an agent by running real past cases through it and comparing the results with what the current version produced. It catches quiet changes in behaviour, such as a rule applied in a different order, that unit tests usually miss.

Retry with backoff

Trying a failed call again after a delay that grows with each attempt, often with a little randomness added. It rides out brief outages and rate limits without hammering the service, and it only works safely when the operation is idempotent.

Rollback

The ability to undo what an automation did, or to switch back to the previous version quickly. Reversible writes, drafts before publishing and versioned prompts are what make a team comfortable letting an agent act at all.

Rules engine

Deterministic code that applies business rules, such as prices, discounts or eligibility, the same way every time. Pairing a rules engine with a language model keeps the model on the messy reading and matching, and keeps the arithmetic explainable.

Shadow mode

Running a new agent alongside the existing process so it proposes outputs while people still do the work and make the decisions. Comparing the two over a few weeks shows where the agent is right, where it is wrong and whether it is ready to act.

Structured output

Asking a model to return data in a fixed shape, such as JSON that matches a schema, instead of free text. It makes the output checkable by code, which is the first step to catching errors before they reach a system of record.

Throughput

The number of items a process finishes in a period, such as orders a day or pull requests a week.

Token

The unit models read and write, roughly three quarters of an English word on average. API prices are quoted per million tokens, usually with output tokens costing several times more than input tokens.

Tool calling

A model's ability to request an action, such as a database query or an API call, by returning a structured request that the surrounding code executes. It is how agents act on real systems, and every tool is a permission to scope carefully.

Touch time

The time someone is actively working on an item. Comparing it with cycle time shows how much of a process is waiting.

Tracing

Recording each step of a run, such as model calls, tool calls and their timings, as a linked sequence you can inspect later. Tracing is the backbone of LLM observability tools.

Work in progress

Items that have started but not finished. Too much of it hides the bottleneck and slows everything down.

Want this in your operation

I build and run production AI agents that take repetitive work off operational teams. Tell me what your team spends too long on.

shurco.aiGuidesFree toolsSolutionsInsightsRSSllms.txt