InsurCore

AI Agents & MCP Glossary

AI Agents & MCP Glossary

Plain-language definitions of AI agents, tool use, and the Model Context Protocol (MCP): agents, tools, orchestration, memory, retrieval, and the protocol that connects AI to external systems.

Terms

169 terms, defined in plain language

Vendor-neutral definitions, grouped by topic. Jump to a section, or browse the whole list. For the rest of the collection, see the glossary hub.

AI agents

AI Agent

Agents

A software system that uses a language model to decide and act over multiple steps toward a goal, typically by calling tools, observing results, and planning its next move rather than producing a single fixed response. Agents differ from one-shot prompts in that control flow is driven by the model's decisions.

Agentic

Agents

Describing software behavior in which a model autonomously decides what actions to take, in what order, and when to stop, instead of following a fixed, pre-scripted sequence. The term emphasizes model-driven control flow over hard-coded logic.

Agent Loop

Agents

The repeating cycle in which an agent sends context to a model, receives an action (such as a tool call), executes it, appends the result to the context, and repeats until a stopping condition is met. It is the core control structure of most agent implementations.

ReAct (Reason + Act)

Agents

A prompting pattern in which a model interleaves reasoning steps with actions, producing a thought, then an action, then observing the result before reasoning again. It lets a model use tools while making its intermediate reasoning explicit.

Autonomy Levels

Agents

An informal scale describing how much an agent acts without human approval, ranging from suggestion-only assistants to systems that plan and execute multi-step tasks unsupervised. Higher autonomy increases capability but also the need for guardrails and oversight.

Goal-Directed Behavior

Agents

Agent behavior oriented toward achieving a specified objective, where intermediate actions are chosen because they are expected to advance that objective. It contrasts with reactive, single-turn responses that do not track a longer-term aim.

Tool-Using Agent

Agents

An agent whose capabilities extend beyond text generation through calling external functions, APIs, or services to gather information or take action in the world. Tools let the model affect systems it cannot reach through text alone.

Conversational Agent

Agents

An agent designed to interact through natural-language dialogue, maintaining context across turns and optionally invoking tools to fulfill requests. Chat assistants are a common example.

Coding Agent

Agents

An agent specialized in software tasks that can read and edit files, run commands, execute tests, and iterate based on results. It typically operates inside a repository or development environment with tools for the file system and shell.

Computer Use

Agents

A capability in which a model controls a computer through screenshots and simulated mouse and keyboard actions, operating graphical interfaces the way a person would. It allows automation of applications that lack a programmatic API.

Browser Agent

Agents

An agent that navigates and acts within a web browser, reading page content and clicking, typing, or extracting data to complete tasks on websites. It may use a rendered-page view, the accessibility tree, or raw HTML.

System Prompt

Prompting

The instructions given to a model that set its role, behavior, constraints, and available context, separate from the user's message. For agents it typically defines the persona, rules, and how tools should be used.

Prompt

Prompting

The input text given to a language model to elicit a response, which may include instructions, examples, context, and a question or task. Its wording and structure strongly influence the output.

Prompt Engineering

Prompting

The practice of designing and refining prompts to get reliable, high-quality model behavior for a task. It covers wording, structure, examples, and the ordering of information.

Few-Shot Prompting

Prompting

Including a handful of examples of the desired input-output behavior in the prompt to guide the model. It can steer format and style without any model fine-tuning.

Zero-Shot Prompting

Prompting

Asking a model to perform a task from instructions alone, with no examples provided. It relies entirely on the model's pretrained ability to generalize.

Chain-of-Thought (CoT)

Reasoning

A prompting approach that encourages a model to produce intermediate reasoning steps before a final answer, which can improve performance on multi-step problems. The reasoning may be shown or kept internal depending on the setup.

Reflection

Reasoning

An agent technique in which the model reviews its own prior output or actions, identifies mistakes, and revises its approach. Reflection loops can improve quality on tasks where self-critique is reliable.

Self-Critique

Reasoning

A step in which a model evaluates its own response against criteria and points out weaknesses before producing or revising a final answer. It is a building block of reflection and evaluator-optimizer patterns.

Reasoning Model

Reasoning

A model trained or configured to spend additional computation on internal step-by-step reasoning before answering, often improving results on hard problems. The extra reasoning may be hidden from the user.

Extended Thinking

Reasoning

A mode in which a model is allowed to generate a longer internal reasoning process before its final response, trading latency and cost for better results on complex tasks. The amount of thinking can often be configured.

Multimodal

Agents

Describing a model or agent that can process more than one kind of input or output, such as text together with images, audio, or documents. Multimodal capability lets agents handle tasks involving visual or other non-text data.

Agent Skill

Agents

A packaged set of instructions, and sometimes scripts or resources, that an agent can load to perform a particular kind of task well. Skills extend an agent's abilities in a reusable, shareable form.

Environment

Agents

The external world an agent perceives and acts upon through its tools, such as a file system, web, database, or application. The environment supplies observations and reflects the effects of the agent's actions.

Action Space

Agents

The full set of actions an agent is permitted to take at a given point, typically defined by the tools available to it. Constraining the action space is a direct way to bound an agent's behavior.

Observation

Agents

The information an agent receives back from the environment after taking an action, such as a tool result or updated state. Observations inform the agent's next decision in the loop.

Long-Horizon Task

Agents

A task that requires many steps or extended time to complete, where the agent must maintain coherence and progress across a long sequence of actions. Such tasks stress memory, planning, and error recovery.

Agent Persona

Prompting

The defined role, tone, and behavioral guidelines an agent adopts, usually established in its system prompt. A consistent persona shapes how the agent communicates and makes decisions.

Research Agent

Agents

An agent that answers complex questions by iteratively searching, reading, and synthesizing information from many sources before producing a report. It typically runs many retrieval and reasoning steps and may use subagents.

Model Context Protocol (MCP)

Model Context Protocol (MCP)

MCP

An open protocol, introduced by Anthropic in late 2024, that standardizes how AI applications connect to external tools and data sources. It defines a common way for AI hosts to discover and use capabilities exposed by servers, so an integration built once can work across many applications.

MCP Server

MCP

A program that exposes capabilities such as tools, resources, and prompts to AI applications over the Model Context Protocol. A server typically wraps a specific system, such as a database, file system, or SaaS API, behind the standard protocol.

MCP Client

MCP

The protocol component, created and managed by a host application, that maintains a one-to-one connection to a single MCP server and exchanges protocol messages with it. Each server a host uses is reached through its own client.

MCP Host

MCP

The AI application that coordinates one or more MCP clients and connects the model to the servers they reach, such as a desktop assistant or IDE. The host decides which servers to connect to and how their capabilities are surfaced to the model and user.

MCP Tools

MCP

An MCP primitive consisting of executable functions a server exposes that the model can invoke to take actions or fetch computed results, each with a name, description, and input schema. Tools are model-controlled: the model decides when to call them.

MCP Resources

MCP

An MCP primitive representing data a server makes available for context, such as files, database records, or documents, each identified by a URI. Resources are application-controlled and typically read-only, intended to supply context rather than perform actions.

MCP Prompts

MCP

An MCP primitive consisting of reusable prompt templates or workflows a server offers, often surfaced to users as commands or starting points. Prompts are user-controlled, meant to be selected deliberately rather than invoked autonomously by the model.

MCP Primitives

MCP

The core building blocks a server can expose over MCP: tools (model-controlled actions), resources (application-controlled context data), and prompts (user-controlled templates). The three differ by who is expected to decide when each is used.

Tools vs. Resources vs. Prompts (MCP)

MCP

In MCP, tools are actions the model chooses to invoke, resources are read-only data the application pulls in for context, and prompts are templates the user deliberately selects. The distinction is about control: model-controlled, application-controlled, and user-controlled respectively.

MCP Transport

MCP

The communication layer that carries MCP messages between a client and server, defining how bytes are sent and received. MCP specifies transports including standard input/output and HTTP-based options.

stdio Transport

MCP

An MCP transport in which the host launches the server as a subprocess and exchanges JSON-RPC messages over its standard input and output streams. It suits local servers running on the same machine as the host.

Streamable HTTP Transport

MCP

An MCP transport in which the client communicates with a server over HTTP, with the server able to stream messages back, optionally using server-sent events. It supports remote servers and replaced the earlier standalone HTTP+SSE transport.

HTTP + SSE Transport

MCP

An earlier MCP transport that used HTTP POST for client-to-server messages and a separate server-sent events stream for server-to-client messages. It has largely been superseded by the streamable HTTP transport.

JSON-RPC

MCP

A lightweight remote-procedure-call protocol that encodes requests, responses, and notifications as JSON objects. MCP uses JSON-RPC 2.0 as the wire format for all messages between clients and servers.

MCP Capabilities

MCP

The features a client and server declare they support during connection setup, such as which primitives are available and whether notifications like list changes are supported. Capability negotiation lets each side adapt to what the other can do.

MCP Initialization

MCP

The handshake at the start of an MCP connection in which the client and server exchange protocol versions and capabilities before normal operation begins. It establishes a shared understanding of what each side supports.

MCP Sampling

MCP

An MCP feature that lets a server request a model completion from the client, so the server can use the host's model without holding its own credentials. The client stays in control, able to review or deny the request.

MCP Roots

MCP

A mechanism by which a client tells a server which file-system or URI locations are in scope for the session. Roots help a server constrain its operations to directories the host has authorized.

MCP Elicitation

MCP

An MCP feature that lets a server ask the user for additional information through the client during an interaction, such as prompting for a missing parameter. It enables servers to gather input without embedding their own user interface.

MCP vs. Function Calling

MCP

Function calling is a model capability for emitting structured tool requests, while MCP is a protocol that standardizes how applications connect to the servers providing those tools and data. MCP complements function calling by making integrations portable across hosts instead of hard-wired into one application.

Tool Discovery (MCP)

MCP

The process by which a client asks a connected server for the list of tools it offers and their schemas, so the host can present them to the model. Discovery lets hosts work with servers without prior knowledge of their specific tools.

List Changed Notification (MCP)

MCP

An MCP message a server sends to tell the client that its set of tools, resources, or prompts has changed, prompting the client to refresh. It keeps the host's view of a server's capabilities current.

Resource URI (MCP)

MCP

The unique identifier a server assigns to each resource so clients can reference and read it. The URI scheme is defined by the server and may point to files, records, or other data.

MCP Registry

MCP

A catalog of available MCP servers that helps hosts and users discover and install integrations. Registries make it easier to find servers for common systems without building connectors from scratch.

Remote MCP Server

MCP

An MCP server reached over a network using an HTTP-based transport rather than run as a local subprocess. Remote servers can be hosted centrally and shared, often with authentication such as OAuth.

Local MCP Server

MCP

An MCP server that runs on the same machine as the host, commonly connected over the stdio transport. Local servers are well suited to accessing local files, applications, or credentials.

MCP Authorization (OAuth)

MCP

The framework MCP defines for authorizing access to remote servers, based on OAuth 2.1, so a host can obtain and present tokens on the user's behalf. It lets servers protect access to sensitive data and actions.

Capability Negotiation

MCP

The exchange at the start of a connection in which two parties agree on which features each supports before using them. In MCP, clients and servers negotiate capabilities during initialization so neither relies on unsupported features.

Connector

MCP

An integration that links an AI application to an external system's data or actions. MCP servers are a standardized form of connector, replacing one-off integrations with a common protocol.

N-times-M Integration Problem

MCP

The challenge that connecting N AI applications to M external systems with bespoke integrations requires building and maintaining up to N times M connectors. A shared protocol like MCP reduces this to building each application and each system to one common standard.

Tool use & function calling

Tool Use

Tools

The general capability of a language model to invoke external functions during generation, receiving structured results it can reason over and continue from. Tool use connects a model to data and actions beyond its training.

Function Calling

Tools

A mechanism in which a model, given descriptions of available functions, outputs a structured request naming a function and its arguments instead of free text. The application executes the function and returns the result for the model to use.

Tool Calling

Tools

The act of a model emitting a request to invoke a named tool with specific arguments during its response. It is the runtime event that drives an agent loop, and 'tool calling' and 'function calling' are often used interchangeably.

Tool Definition

Tools

The declaration a model is given for a tool, including its name, a natural-language description, and a schema for its inputs. Clear, unambiguous definitions strongly affect how well a model chooses and calls the tool.

Tool Result

Tools

The output returned to the model after a tool is executed, appended to the conversation so the model can reason over it and continue. Results may be success values, errors, or structured data.

Tool Schema

Tools

The formal specification, commonly JSON Schema, describing a tool's input parameters, their types, and which are required. The model uses it to produce validly structured arguments.

JSON Schema

Schema

A standard vocabulary for describing the structure of JSON data, including field names, types, constraints, and required properties. It is widely used to define tool inputs and to validate structured model outputs.

Structured Output

Output

Model output constrained to a defined format such as JSON matching a schema, rather than free-form prose. It makes responses reliably machine-parsable for downstream code.

Constrained Decoding

Output

A generation technique that restricts the tokens a model may emit at each step so the output conforms to a grammar or schema. It can guarantee syntactically valid structured output such as well-formed JSON.

Parallel Tool Calls

Tools

A single model turn that requests more than one tool invocation at once, allowing independent calls to execute concurrently. It reduces round trips when several pieces of information are needed together.

Tool Choice

Tools

A setting that controls whether and how a model calls tools, such as letting it decide automatically, forcing it to call any tool, or requiring a specific named tool. It gives developers control over tool-use behavior.

Tool Description

Tools

The natural-language explanation of what a tool does, when to use it, and what it returns, provided as part of a tool definition. Because the model relies on it to decide usage, precise descriptions improve tool selection.

Tool Error Handling

Tools

The practice of returning informative errors from tools so the model can recover, retry, or choose a different approach rather than failing silently. Well-formed error results help an agent self-correct.

API Tool

Tools

A tool that wraps a call to an external web API, translating the model's structured request into an HTTP request and returning the response. It is a common way to give agents access to live services.

Code Execution Tool

Tools

A tool that runs code the model produces, typically in a sandboxed environment, and returns the output. It lets a model compute, transform data, or verify results rather than reasoning about them in text alone.

Retrieval Tool

Tools

A tool that searches a knowledge source, such as a document store or vector index, and returns relevant passages for the model to use. It is the mechanism by which agents ground answers in external data.

Structured Generation

Output

Producing output that conforms to a required format or schema throughout generation, so the result is directly usable by code. It encompasses techniques like schema-guided decoding and strict JSON output.

Schema Validation

Schema

Checking that data, such as a model's tool arguments or structured output, conforms to a defined schema before it is used. Validation catches malformed values early and lets the system reject or repair them.

Tool Overload

Tools

The degradation in tool selection that can occur when an agent is given too many tools or overlapping ones, making it harder for the model to choose correctly. Curating a focused tool set often improves reliability.

Dynamic Tool Loading

Tools

Exposing only the tools relevant to the current context to the model, rather than all tools at once, to keep selection accurate and context small. Tools can be added or removed as the task changes.

Output Parsing

Output

Converting a model's textual response into a structured form the application can act on, such as extracting fields from JSON. Reliable parsing often depends on the model producing well-formed structured output.

Orchestration & workflows

Planner

Orchestration

The component or model step that decomposes a goal into an ordered set of subtasks or actions before or during execution. Planning may happen once up front or be revised continuously as new information arrives.

Plan-and-Execute

Orchestration

An agent pattern that first produces a full multi-step plan, then carries out each step, optionally re-planning if steps fail. Separating planning from execution can improve reliability on complex tasks compared with deciding one step at a time.

Orchestration

Orchestration

The coordination of models, tools, and control flow so that a multi-step task runs in the right order with the right inputs. Orchestration logic decides when to call a model, when to call a tool, and how to route intermediate results.

Orchestrator-Worker Pattern

Multi-agent

A multi-agent design in which a central orchestrator breaks a task into pieces and delegates each to worker agents, then combines their outputs. It suits tasks that split cleanly into parallelizable or specialized subtasks.

Workflow

Workflows

A system in which models and tools are connected through predefined code paths, so the sequence of steps is largely fixed by the developer rather than decided by the model. Workflows trade flexibility for predictability and are often preferable when the steps are known in advance.

Agent vs. Workflow

Workflows

A distinction between systems where the model dynamically directs its own process and tool use (agents) and systems where models follow fixed, developer-defined paths (workflows). Agents offer flexibility for open-ended tasks; workflows offer consistency for well-understood ones.

Prompt Chaining

Workflows

A workflow pattern that decomposes a task into a fixed sequence of model calls, where each call's output feeds the next. It improves accuracy on multi-part tasks at the cost of added latency.

Routing

Workflows

A workflow pattern that classifies an input and directs it to a specialized prompt, model, or tool suited to that category. Routing lets each path be optimized independently instead of forcing one prompt to handle every case.

Parallelization

Workflows

Running multiple model or tool calls at the same time, either by splitting a task into independent subtasks (sectioning) or by running the same task several times and aggregating (voting). It reduces latency and can improve quality through aggregation.

Evaluator-Optimizer

Workflows

A workflow in which one model generates a candidate response and another evaluates it and gives feedback, looping until the output meets a standard. It works well when clear evaluation criteria exist and iterative refinement adds value.

Multi-Agent System

Multi-agent

A system composed of several agents that collaborate, each often with a distinct role, tool set, or specialization, to solve a task larger than any one agent handles well. Coordination between agents becomes a central design concern.

Subagent

Multi-agent

An agent invoked by another agent to handle a scoped subtask, typically with its own context, tools, and instructions, returning only a result to the caller. Subagents keep specialized work out of the parent's context window.

Handoff

Multi-agent

The transfer of control of a task from one agent to another, passing along the context the receiving agent needs to continue. Handoffs let specialized agents take over the portions of a task they are best suited to.

Delegation

Multi-agent

The act of an agent assigning part of its work to a subagent or tool rather than performing it directly. Effective delegation depends on a clear brief, since the delegate sees only the context it is given.

Supervisor Agent

Multi-agent

An agent that manages other agents, deciding which to invoke, reviewing their outputs, and synthesizing a final result. It acts as the coordination layer in a multi-agent system.

Stop Condition

Orchestration

The criterion that ends an agent loop, such as the model signaling completion, reaching a step or token limit, or satisfying a goal check. Well-chosen stop conditions prevent both premature halting and runaway loops.

Task Decomposition

Orchestration

Breaking a complex goal into smaller, more manageable subtasks that can be solved individually and recombined. It is a core strategy enabling planners and multi-agent systems to handle large problems.

Deterministic vs. Probabilistic

Workflows

A contrast between code paths that always behave the same way and model outputs that can vary between runs. Robust agent systems place deterministic checks around probabilistic model decisions to ensure reliability.

Context, memory & retrieval

Context Window

Context

The maximum amount of text, measured in tokens, that a model can consider at once, including the prompt, prior conversation, tool results, and its own output. Everything the model reasons over for a turn must fit within this limit.

Token

Context

The basic unit of text a language model processes, often a word fragment rather than a whole word. Context length, cost, and latency are typically measured in tokens.

Token Budget

Context

The portion of a context window allocated to a given purpose, such as instructions, retrieved documents, or history, within the overall limit. Managing the budget keeps the most relevant information in context and controls cost.

Context Engineering

Context

The practice of deciding what information to place in a model's context window, and how to structure it, so the model has what it needs without being overloaded. It extends prompt engineering to cover tool results, memory, and retrieved data.

Context Degradation

Context

The tendency for a model's reliability to decline as its context fills with large amounts of text, as relevant details get diluted among less relevant ones. It motivates keeping context focused rather than maximally full.

Short-Term Memory

Memory

The information an agent holds for the current task or session, typically the running conversation and recent tool results kept within the context window. It is lost or must be summarized once the session ends or the window fills.

Long-Term Memory

Memory

Information an agent persists across sessions, stored outside the context window and retrieved when relevant, such as user preferences or facts learned earlier. It gives an agent continuity beyond a single conversation.

Episodic Memory

Memory

A form of agent memory that records specific past interactions or events, so the agent can recall what happened in a prior session. It contrasts with semantic memory, which stores general facts rather than particular episodes.

Semantic Memory

Memory

A form of agent memory that stores general facts and knowledge independent of when they were learned, such as a user's role or a project's conventions. It is drawn on to inform behavior across many tasks.

Memory Store

Memory

The external storage, such as a database or file, where an agent writes and reads persistent memory. It lets information survive beyond the context window and be recalled in later sessions.

Scratchpad

Memory

A space, often within the context or an external note, where an agent records intermediate reasoning, plans, or partial results as it works. It helps the agent track progress on multi-step tasks.

Conversation History

Context

The accumulated record of prior turns in a dialogue that is supplied to the model so it can maintain continuity. As it grows it consumes context and may need truncation or summarization.

Context Summarization

Context

The technique of compressing earlier conversation or tool output into a shorter summary so the essential information fits within the context window. It trades some detail for the ability to continue long interactions.

Retrieval-Augmented Generation (RAG)

Retrieval

A technique that retrieves relevant documents from an external source and includes them in the model's context so answers are grounded in that data. It lets a model use up-to-date or private information without being retrained.

Retrieval

Retrieval

The process of finding and returning information relevant to a query from a larger collection, for use as model context. It underlies RAG and many agent tools that search knowledge sources.

Embedding

Retrieval

A numeric vector that represents the meaning of a piece of text, such that similar meanings map to nearby vectors. Embeddings enable semantic search by comparing vector distances rather than matching exact words.

Vector Store

Retrieval

A database optimized for storing embeddings and finding those most similar to a query vector. It is the backbone of semantic retrieval in many RAG and agent systems.

Chunking

Retrieval

Splitting documents into smaller passages before embedding and indexing, so retrieval returns focused, relevant pieces rather than whole files. Chunk size and overlap affect retrieval quality.

Reranking

Retrieval

A second-stage step that reorders an initial set of retrieved candidates using a more precise relevance model before passing the best to the model. It improves the quality of the passages that reach the context window.

Grounding

Retrieval

Tying a model's output to verifiable external information, such as retrieved documents or tool results, so claims can be traced to a source. Grounding reduces unsupported statements and supports citation.

Knowledge Base

Retrieval

A curated collection of documents or structured data an agent can retrieve from to answer questions or inform actions. It is the source grounded responses draw on.

Citation

Retrieval

A reference an agent includes that points to the source supporting a statement, such as a specific retrieved passage. Citations let users verify grounded claims and build trust in the output.

Memory Retrieval

Memory

Fetching relevant items from an agent's long-term memory store and placing them in context when they bear on the current task. It lets persistent memory inform behavior without overwhelming the context window.

Result Truncation

Context

Shortening a large tool output before adding it to the context so it does not consume excessive tokens, for example by returning a summary or the first portion with a pointer to the rest. It keeps context focused on what the model needs.

Progressive Disclosure

Context

A context strategy in which an agent is given high-level pointers and loads detailed information only when a task requires it, rather than all details up front. It keeps the context window lean while preserving access to depth.

Evaluation & guardrails

Trajectory

Evaluation

The full sequence of states, actions, tool calls, and observations an agent produces while working through a task. Trajectories are examined to debug behavior and evaluate agent performance.

Guardrails

Safety

Controls that constrain an agent's behavior to keep it safe, on-task, and compliant, such as input and output filters, allowed-tool lists, and approval gates. They reduce the risk of harmful or unintended actions.

Human-in-the-Loop (HITL)

Safety

A design in which a person reviews or approves certain agent actions before they take effect, especially consequential or irreversible ones. It balances automation with human oversight.

Human-on-the-Loop

Safety

A supervision model in which a person monitors an agent operating autonomously and can intervene when needed, rather than approving each action in advance. It suits higher-throughput automation with retained oversight.

Approval Gate

Safety

A checkpoint where an agent must pause for explicit authorization before performing a specified action, such as sending an email or deleting data. Gates are a concrete way to implement human-in-the-loop control.

Evaluation (Eval)

Evaluation

A structured test that measures how well a model or agent performs on a task, using defined inputs and scoring criteria. Evals are used to compare approaches and catch regressions before deployment.

LLM-as-a-Judge

Evaluation

An evaluation method that uses a language model to score or compare outputs against criteria, in place of or alongside human raters. It scales evaluation but can inherit the judging model's biases and must be validated.

Benchmark

Evaluation

A standardized dataset and scoring method used to measure and compare model or agent capabilities on a particular kind of task. Benchmarks enable comparison but may not reflect real-world performance.

Golden Dataset

Evaluation

A curated set of inputs paired with known-correct outputs, used as ground truth when evaluating a model or agent. It provides a stable reference for measuring accuracy and detecting regressions.

Regression Testing

Evaluation

Re-running a fixed set of evals after a change to confirm that previously working behavior has not degraded. It guards against silent quality drops when prompts, models, or tools change.

Hallucination

Safety

A model output that is fluent and confident but factually wrong or unsupported by its sources. Grounding, retrieval, and verification steps are common ways to reduce it.

Prompt Injection

Security

An attack in which malicious instructions are placed in content the model reads, such as a web page, document, or tool result, to hijack its behavior. It is a core risk for agents that process untrusted data.

Indirect Prompt Injection

Security

A prompt injection delivered through external data an agent later retrieves, rather than typed by the user, such as hidden text in a fetched document. Because the agent trusts tool output, such instructions can be especially hard to spot.

Jailbreak

Security

An attempt to bypass a model's safety guidelines or restrictions through crafted prompts, so it produces content or takes actions it is meant to refuse. Jailbreaks exploit gaps between stated rules and model behavior.

Data Exfiltration

Security

The unauthorized extraction of sensitive data from a system, which an agent can be tricked into performing, for example by a prompt injection that directs it to send private information outward. It is a key threat in agent security.

Least Privilege

Security

The principle of granting an agent only the minimum tools, data access, and permissions needed for its task. Limiting scope reduces the damage possible if the agent misbehaves or is compromised.

Content Filter

Safety

A classifier that screens model inputs or outputs for disallowed content and blocks or flags it. Filters are a common guardrail against unsafe or policy-violating text.

Red Teaming

Security

The practice of deliberately probing a model or agent with adversarial inputs to find safety and security weaknesses before attackers do. Findings inform guardrails and fixes.

Loop Guard

Safety

A safeguard that caps the number of iterations or repeated actions an agent may take, preventing infinite loops when it fails to make progress. It bounds cost and protects against stuck behavior.

Verification Step

Evaluation

An explicit check an agent performs to confirm a result is correct before proceeding or finishing, such as running tests or re-querying a source. Verification catches errors that self-reported confidence would miss.

PII Redaction

Security

Removing or masking personally identifiable information from data before it is sent to a model or logged. Redaction protects sensitive data and supports privacy and compliance requirements.

Agent infrastructure

Sandbox

Infra

An isolated environment in which an agent's code execution or actions are confined so they cannot affect the host system or production data. Sandboxing limits the blast radius of mistakes and attacks.

Observability

Infra

The ability to inspect what an agent did, through logs, traces, and metrics of its prompts, tool calls, and outputs. Good observability is essential for debugging and monitoring agent systems.

Tracing

Infra

Recording the detailed sequence of steps in an agent run, including each model call, tool invocation, inputs, and outputs, usually linked by a shared identifier. Traces let developers follow and diagnose a specific execution.

Streaming

Infra

Delivering a model's output incrementally as it is generated rather than all at once when complete. Streaming lowers perceived latency and lets a client display or act on partial results.

Rate Limiting

Infra

Capping how many requests or tokens may be sent to a model or tool in a time window, to control cost and protect shared services. Agents must handle rate-limit responses, often by backing off and retrying.

Retry with Backoff

Infra

An error-handling strategy that reattempts a failed call after progressively longer waits, often with randomized jitter. It helps agents recover from transient failures and rate limits without overwhelming a service.

Idempotency

Infra

A property of an operation that produces the same result whether performed once or repeated, so retries do not cause duplicate effects. It is important when agents may re-execute actions after failures.

Prompt Caching

Infra

Reusing the processed form of a repeated prefix of a prompt across requests so the model does not reprocess it each time. It reduces latency and cost when a large, stable context is sent on many calls.

Token Usage

Infra

The count of input and output tokens consumed by a model call, which typically determines its cost and counts against the context limit. Tracking usage is central to managing agent cost and performance.

Latency

Infra

The time between sending a request to a model or tool and receiving a usable response. In multi-step agents, latency accumulates across turns and is a key factor in user experience.

Temperature

Infra

A sampling parameter that controls the randomness of a model's output, with lower values producing more deterministic, focused responses and higher values more varied ones. Agents often use low temperature for reliable tool use.

Max Tokens

Infra

A limit on how many tokens a model may generate in a single response. Setting it controls output length, cost, and the risk of truncated responses.

State Management

Infra

How an agent system tracks and persists the evolving state of a task, including progress, intermediate results, and pending actions, across steps and possibly restarts. Durable state enables resuming long-running work.

Checkpointing

Infra

Periodically saving an agent's state so a long-running task can resume from the last saved point after an interruption. It adds resilience to multi-step and long-horizon work.

Fine-Tuning

Infra

Further training a pretrained model on task- or domain-specific examples to adjust its behavior. It is one way to specialize a model, as an alternative or complement to prompting and retrieval.

Inference

Infra

The process of running a trained model to produce output from an input, as opposed to training the model. Each agent step that calls the model is an inference request.

Model Router

Infra

A component that selects which model to use for a given request, often choosing a smaller, cheaper model for easy tasks and a larger one for hard tasks. Routing can cut cost and latency while preserving quality.

Agent Framework

Infra

A software library or platform that provides building blocks for agents, such as loops, tool integration, memory, and orchestration, so developers do not implement them from scratch. Frameworks vary in how much control versus abstraction they offer.

Tool Registry

Infra

A catalog of the tools available to an agent, holding their definitions and schemas, from which the host selects which to expose to the model. It centralizes tool management across an application.

Credential Management

Security

The secure handling of the secrets, such as API keys and tokens, that tools and servers need to access protected systems. Proper management keeps secrets out of prompts, logs, and model output.

Cost Tracking

Infra

Monitoring the spending an agent incurs from model and tool calls, usually derived from token usage and per-call pricing. It supports budgeting and catching runaway or inefficient behavior.

Tool Result Caching

Infra

Storing the result of a tool call so that an identical later call can reuse it instead of re-executing. It saves cost and latency for repeated or deterministic lookups within or across agent runs.

Bring Us the Backup. We Will Help Carry the Conversion Across the Finish Line.

Whether you need overflow capacity for one acquisition or a repeatable conversion partner across an active pipeline, we'll help you define the scope, inspect the source data, and build a practical path into Applied Epic.

Built for insurance agencies, brokerages, aggregators, and acquisition teams that need experienced conversion capacity without building another internal department.

Schedule an Applied Epic Conversion Review

No commitment required. We typically respond within one business day.