# AI dictionary for Fast-Moving Agent Terms | Joel Saji

> A plain-English AI dictionary for fast-moving terms: agents, long-horizon tasks, harnesses, MCP, RAG, evals, and guardrails, explained for builders.

Canonical: https://www.joelcsaji.com/ai-dictionary

145 entries, each with its own page (definition, example, related terms, and sources). The JSON API is at https://www.joelcsaji.com/api/ai-terms/{slug}.

- [AI Agent](https://www.joelcsaji.com/ai-dictionary/ai-agent) (`ai-agent`, Agents & Autonomy, Established): A system that uses a model, tools, context, and control logic to pursue a goal and take actions.
- [Agentic AI](https://www.joelcsaji.com/ai-dictionary/agentic-ai) (`agentic-ai`, Agents & Autonomy, Established): AI that plans, uses tools, receives feedback from the environment, and can act with partial autonomy.
- [Agent Harness](https://www.joelcsaji.com/ai-dictionary/agent-harness) (`agent-harness`, Agents & Autonomy, Rising): The runtime around a model that supplies tools, state, prompts, permissions, logs, and execution loops.
- [Harness Engineering](https://www.joelcsaji.com/ai-dictionary/harness-engineering) (`harness-engineering`, Agents & Autonomy, Rising): The practice of designing, testing, and simplifying the runtime around an agent so it can work reliably across tools, sessions, and tasks.
- [Brain-Hands Decoupling](https://www.joelcsaji.com/ai-dictionary/brain-hands-decoupling) (`brain-hands-decoupling`, Agents & Autonomy, Emerging): Separating the model-and-harness layer that decides what to do from the sandbox, tools, or systems that perform actions.
- [Agent Loop](https://www.joelcsaji.com/ai-dictionary/agent-loop) (`agent-loop`, Agents & Autonomy, Established): The repeated cycle where an agent reasons, calls a tool, observes the result, and decides the next step.
- [Long-Horizon Task](https://www.joelcsaji.com/ai-dictionary/long-horizon-task) (`long-horizon-task`, Agents & Autonomy, Rising): A goal that requires many dependent steps, decisions, tool calls, and course corrections before completion.
- [Long-Running Agent](https://www.joelcsaji.com/ai-dictionary/long-running-agent) (`long-running-agent`, Agents & Autonomy, Emerging): An agent designed to keep making progress over many sessions, context windows, or wall-clock hours.
- [Background Mode](https://www.joelcsaji.com/ai-dictionary/background-mode) (`background-mode`, Agents & Autonomy, Rising): An execution mode where a long-running AI task continues asynchronously while the application polls for status or results.
- [Agent Swarm](https://www.joelcsaji.com/ai-dictionary/agent-swarm) (`agent-swarm`, Agents & Autonomy, Emerging): A group of agents that coordinate or compete to solve pieces of a larger goal.
- [Goal Mode](https://www.joelcsaji.com/ai-dictionary/goal-mode) (`goal-mode`, Agents & Autonomy, Watch): An emerging coding-agent workflow where a persistent objective stays active across progress checks, continuation, pause, and resume.
- [Durable Execution](https://www.joelcsaji.com/ai-dictionary/durable-execution) (`durable-execution`, Agents & Autonomy, Rising): Saving workflow progress so an agent or graph can pause, fail, or wait for a person and later resume without redoing completed steps.
- [Interrupt](https://www.joelcsaji.com/ai-dictionary/interrupt) (`interrupt`, Agents & Autonomy, Rising): A deliberate pause inside an agent workflow that waits for external input before continuing.
- [Multi-Agent System](https://www.joelcsaji.com/ai-dictionary/multi-agent-system) (`multi-agent-system`, Agents & Autonomy, Rising): An architecture where multiple specialized agents collaborate through messages, tools, or a shared controller.
- [Supervisor Agent](https://www.joelcsaji.com/ai-dictionary/supervisor-agent) (`supervisor-agent`, Agents & Autonomy, Rising): A coordinating agent that decomposes work, invokes specialists, and combines their outputs.
- [Sub-Agent](https://www.joelcsaji.com/ai-dictionary/sub-agent) (`sub-agent`, Agents & Autonomy, Rising): A focused agent invoked by another agent or workflow to handle a narrower task.
- [Agent-as-Tool](https://www.joelcsaji.com/ai-dictionary/agent-as-tool) (`agent-as-tool`, Agents & Autonomy, Emerging): A pattern where one agent exposes a specialist agent as a callable tool instead of handing over the whole conversation.
- [Agent Team](https://www.joelcsaji.com/ai-dictionary/agent-team) (`agent-team`, Agents & Autonomy, Emerging): Multiple agent sessions coordinated around shared work, often with peer messaging, task assignment, or separate context.
- [Handoff](https://www.joelcsaji.com/ai-dictionary/handoff) (`handoff`, Agents & Autonomy, Rising): A controlled transfer of a task or conversation from one agent to another.
- [Orchestrator-Workers](https://www.joelcsaji.com/ai-dictionary/orchestrator-workers) (`orchestrator-workers`, Agents & Autonomy, Established): A pattern where a central model dynamically breaks work into subtasks and delegates them to worker models.
- [Evaluator-Optimizer](https://www.joelcsaji.com/ai-dictionary/evaluator-optimizer) (`evaluator-optimizer`, Agents & Autonomy, Established): A workflow where one model produces work and another model evaluates it, creating an improvement loop.
- [Human-in-the-Loop](https://www.joelcsaji.com/ai-dictionary/human-in-the-loop) (`human-in-the-loop`, Agents & Autonomy, Established): A design where a person reviews, approves, or edits model output before a consequential action happens.
- [Human-on-the-Loop](https://www.joelcsaji.com/ai-dictionary/human-on-the-loop) (`human-on-the-loop`, Agents & Autonomy, Rising): A design where humans monitor autonomous systems and intervene when risk, uncertainty, or policy requires it.
- [Autonomy Level](https://www.joelcsaji.com/ai-dictionary/autonomy-level) (`autonomy-level`, Agents & Autonomy, Emerging): A way to describe how much freedom an agent has to decide, act, spend money, or change systems.
- [Context Window](https://www.joelcsaji.com/ai-dictionary/context-window) (`context-window`, Context & Memory, Established): The amount of text, images, tool results, or other tokens a model can consider in one request.
- [Token](https://www.joelcsaji.com/ai-dictionary/token) (`token`, Context & Memory, Established): A chunk of text or data that a model reads or writes while processing a prompt or response.
- [Prompt](https://www.joelcsaji.com/ai-dictionary/prompt) (`prompt`, Context & Memory, Established): The text, media, files, or structured messages sent to a model to guide its response.
- [Prompt Engineering](https://www.joelcsaji.com/ai-dictionary/prompt-engineering) (`prompt-engineering`, Context & Memory, Established): The practice of writing and testing instructions so a model produces useful, consistent outputs.
- [Developer Message](https://www.joelcsaji.com/ai-dictionary/developer-message) (`developer-message`, Context & Memory, Rising): A higher-priority instruction from the application that sets rules, tone, business logic, or boundaries for the model.
- [Context Truncation](https://www.joelcsaji.com/ai-dictionary/context-truncation) (`context-truncation`, Context & Memory, Established): Removing or clipping older or excess input when a prompt grows beyond the model's context limit.
- [Context Engineering](https://www.joelcsaji.com/ai-dictionary/context-engineering) (`context-engineering`, Context & Memory, Rising): The practice of selecting, ordering, compressing, and updating the information a model sees.
- [Memory](https://www.joelcsaji.com/ai-dictionary/memory) (`memory`, Context & Memory, Rising): Information stored outside the immediate prompt so an AI system can recall preferences, facts, or progress later.
- [Episodic Memory](https://www.joelcsaji.com/ai-dictionary/episodic-memory) (`episodic-memory`, Context & Memory, Emerging): Memory organized around events, interactions, runs, or user sessions.
- [Semantic Memory](https://www.joelcsaji.com/ai-dictionary/semantic-memory) (`semantic-memory`, Context & Memory, Emerging): Memory organized around durable facts, user preferences, and reusable knowledge.
- [Compaction](https://www.joelcsaji.com/ai-dictionary/compaction) (`compaction`, Context & Memory, Rising): Compressing earlier context into a shorter summary so an agent can continue without exceeding its context window.
- [Context Reset](https://www.joelcsaji.com/ai-dictionary/context-reset) (`context-reset`, Context & Memory, Rising): Clearing an agent's active context window and starting a fresh run from a structured handoff artifact.
- [Least-Privilege Context](https://www.joelcsaji.com/ai-dictionary/least-privilege-context) (`least-privilege-context`, Context & Memory, Emerging): A context engineering pattern that gives each agent only the information, tools, and state needed for its role or current task.
- [Checkpoint](https://www.joelcsaji.com/ai-dictionary/checkpoint) (`checkpoint`, Context & Memory, Rising): A durable record of progress, decisions, files, or state that lets an agent resume safely.
- [Prompt Caching](https://www.joelcsaji.com/ai-dictionary/prompt-caching) (`prompt-caching`, Context & Memory, Established): Reusing unchanged prompt or context segments so repeated model calls can be cheaper or faster.
- [Agent Instructions File](https://www.joelcsaji.com/ai-dictionary/agent-instructions-file) (`agent-instructions-file`, Context & Memory, Rising): A repository or workspace file (often AGENTS.md) that gives coding agents durable project rules, commands, conventions, and boundaries.
- [Tool Use](https://www.joelcsaji.com/ai-dictionary/tool-use) (`tool-use`, Tools & Protocols, Established): Letting a model request external functions, APIs, databases, browsers, or code execution.
- [Function Calling](https://www.joelcsaji.com/ai-dictionary/function-calling) (`function-calling`, Tools & Protocols, Established): A tool-use pattern where the model emits structured arguments for an application-defined function.
- [Tool Result](https://www.joelcsaji.com/ai-dictionary/tool-result) (`tool-result`, Tools & Protocols, Established): The data returned to a model after an external tool call completes.
- [Tool Filtering](https://www.joelcsaji.com/ai-dictionary/tool-filtering) (`tool-filtering`, Tools & Protocols, Rising): Selecting a small, relevant subset of tools for an agent instead of loading every available tool definition into context.
- [Agent Identity](https://www.joelcsaji.com/ai-dictionary/agent-identity) (`agent-identity`, Tools & Protocols, Rising): A way to identify an AI agent as an actor with its own metadata, credentials, permissions, and delegation record.
- [Capability Token](https://www.joelcsaji.com/ai-dictionary/capability-token) (`capability-token`, Tools & Protocols, Emerging): A credential that encodes exactly what an agent may do, often including allowed tools, targets, budgets, expiry, and delegation limits.
- [Slash Command](https://www.joelcsaji.com/ai-dictionary/slash-command) (`slash-command`, Tools & Protocols, Rising): A command typed with a leading slash to trigger a tool-specific workflow, setting, skill, or mode before the model responds.
- [Agent Skill](https://www.joelcsaji.com/ai-dictionary/agent-skill) (`agent-skill`, Tools & Protocols, Rising): A packaged set of instructions, reference files, templates, or scripts that an agent can load on demand for a specific capability (often via an open Agent Skills standard).
- [Agent Hook](https://www.joelcsaji.com/ai-dictionary/agent-hook) (`agent-hook`, Tools & Protocols, Emerging): A script, request, prompt, or subagent triggered by an event in an agent's lifecycle.
- [Model Context Protocol](https://www.joelcsaji.com/ai-dictionary/model-context-protocol) (`model-context-protocol`, Tools & Protocols, Established): An open protocol that standardizes how LLM applications connect to tools, data, and contextual resources.
- [MCP Server](https://www.joelcsaji.com/ai-dictionary/mcp-server) (`mcp-server`, Tools & Protocols, Rising): A service that exposes tools, prompts, or resources to an MCP-compatible AI host.
- [MCP Client](https://www.joelcsaji.com/ai-dictionary/mcp-client) (`mcp-client`, Tools & Protocols, Rising): The part of an AI application that connects to MCP servers and presents their capabilities to the model.
- [MCP Resource](https://www.joelcsaji.com/ai-dictionary/mcp-resource) (`mcp-resource`, Tools & Protocols, Rising): Data exposed by an MCP server for a client or model to use as context, identified by a URI.
- [MCP Elicitation](https://www.joelcsaji.com/ai-dictionary/mcp-elicitation) (`mcp-elicitation`, Tools & Protocols, Emerging): An MCP pattern where a server asks the client to gather structured user input or send the user through an out-of-band URL flow.
- [MCP Tasks](https://www.joelcsaji.com/ai-dictionary/mcp-tasks) (`mcp-tasks`, Tools & Protocols, Emerging): An MCP extension that lets long-running tool calls return a task handle, then report progress, request input, or finish later.
- [MCP Apps](https://www.joelcsaji.com/ai-dictionary/mcp-apps) (`mcp-apps`, Tools & Protocols, Emerging): An MCP extension for rendering interactive HTML interfaces, such as forms, dashboards, and visualizations, inside an MCP host.
- [Agent2Agent](https://www.joelcsaji.com/ai-dictionary/agent2agent) (`agent2agent`, Tools & Protocols, Emerging): A protocol for communication between independent agents, especially across services, teams, or frameworks.
- [Agent Card](https://www.joelcsaji.com/ai-dictionary/agent-card) (`agent-card`, Tools & Protocols, Emerging): A machine-readable A2A document that describes an agent's identity, capabilities, interfaces, security, and skills.
- [A2A Task](https://www.joelcsaji.com/ai-dictionary/a2a-task) (`a2a-task`, Tools & Protocols, Emerging): A stateful unit of work exchanged between agents in the Agent2Agent protocol.
- [Artifact](https://www.joelcsaji.com/ai-dictionary/artifact) (`artifact`, Tools & Protocols, Rising): A durable output produced during an agent task, such as a file, report, structured payload, or generated asset.
- [Structured Outputs](https://www.joelcsaji.com/ai-dictionary/structured-outputs) (`structured-outputs`, Tools & Protocols, Established): Constraining model responses to match a defined schema such as JSON with required fields.
- [Streaming](https://www.joelcsaji.com/ai-dictionary/streaming) (`streaming`, Tools & Protocols, Established): Sending model output incrementally as it is generated instead of waiting for the full response.
- [Computer Use](https://www.joelcsaji.com/ai-dictionary/computer-use) (`computer-use`, Tools & Protocols, Rising): A model capability or tool setup that lets an agent see and operate user interfaces with mouse, keyboard, and screenshots.
- [Remote MCP Server](https://www.joelcsaji.com/ai-dictionary/remote-mcp-server) (`remote-mcp-server`, Tools & Protocols, Emerging): An MCP server reachable over the network that exposes external data or tools to an AI application.
- [Retrieval-Augmented Generation](https://www.joelcsaji.com/ai-dictionary/retrieval-augmented-generation) (`retrieval-augmented-generation`, Retrieval & Grounding, Established): A pattern that retrieves external knowledge and adds it to the prompt before generation.
- [Agentic RAG](https://www.joelcsaji.com/ai-dictionary/agentic-rag) (`agentic-rag`, Retrieval & Grounding, Rising): RAG where an agent actively plans searches, chooses sources, reformulates queries, and decides when it has enough evidence.
- [Grounding](https://www.joelcsaji.com/ai-dictionary/grounding) (`grounding`, Retrieval & Grounding, Established): Connecting model output to verifiable sources or environmental facts.
- [Embedding](https://www.joelcsaji.com/ai-dictionary/embedding) (`embedding`, Retrieval & Grounding, Established): A numerical representation of data that captures semantic similarity.
- [Vector Database](https://www.joelcsaji.com/ai-dictionary/vector-database) (`vector-database`, Retrieval & Grounding, Established): A database optimized for storing embeddings and finding similar items quickly.
- [Chunking](https://www.joelcsaji.com/ai-dictionary/chunking) (`chunking`, Retrieval & Grounding, Established): Splitting documents into smaller pieces before embedding, indexing, or sending them to a model.
- [Hybrid Search](https://www.joelcsaji.com/ai-dictionary/hybrid-search) (`hybrid-search`, Retrieval & Grounding, Rising): Combining traditional keyword search with semantic vector search.
- [Reranking](https://www.joelcsaji.com/ai-dictionary/reranking) (`reranking`, Retrieval & Grounding, Established): A second pass that reorders retrieved results based on relevance to the query.
- [Citations](https://www.joelcsaji.com/ai-dictionary/citations) (`citations`, Retrieval & Grounding, Established): Links or references that show which source material supports a model's answer.
- [Reasoning Model](https://www.joelcsaji.com/ai-dictionary/reasoning-model) (`reasoning-model`, Reasoning & Inference, Rising): A model optimized to spend more inference effort on multi-step reasoning, math, code, or planning.
- [Latency](https://www.joelcsaji.com/ai-dictionary/latency) (`latency`, Reasoning & Inference, Established): How long an AI system takes to begin or finish a response, tool call, or workflow.
- [Reasoning Effort](https://www.joelcsaji.com/ai-dictionary/reasoning-effort) (`reasoning-effort`, Reasoning & Inference, Rising): A model setting that trades latency and cost for more or less internal reasoning before producing an answer.
- [Reasoning Summary](https://www.joelcsaji.com/ai-dictionary/reasoning-summary) (`reasoning-summary`, Reasoning & Inference, Rising): A user- or developer-facing summary of a model's reasoning process without exposing the raw hidden chain of thought.
- [Test-Time Compute](https://www.joelcsaji.com/ai-dictionary/test-time-compute) (`test-time-compute`, Reasoning & Inference, Emerging): Spending more compute during inference to improve answer quality without retraining the model.
- [Chain-of-Thought](https://www.joelcsaji.com/ai-dictionary/chain-of-thought) (`chain-of-thought`, Reasoning & Inference, Established): A step-by-step reasoning process used internally or externally to improve complex problem solving.
- [Self-Reflection](https://www.joelcsaji.com/ai-dictionary/self-reflection) (`self-reflection`, Reasoning & Inference, Rising): A model-generated critique or review step intended to improve a prior answer or plan.
- [Model Routing](https://www.joelcsaji.com/ai-dictionary/model-routing) (`model-routing`, Reasoning & Inference, Rising): Choosing different models or workflows based on task type, difficulty, cost, or risk.
- [Parallelization](https://www.joelcsaji.com/ai-dictionary/parallelization) (`parallelization`, Reasoning & Inference, Established): Running multiple model calls or subtasks at the same time and combining their outputs.
- [Constrained Decoding](https://www.joelcsaji.com/ai-dictionary/constrained-decoding) (`constrained-decoding`, Reasoning & Inference, Rising): Restricting token generation so output follows a grammar, schema, or allowed set of values.
- [Sampling Parameters](https://www.joelcsaji.com/ai-dictionary/sampling-parameters) (`sampling-parameters`, Reasoning & Inference, Established): Settings that influence how predictable or varied a model's generated tokens are.
- [Evals](https://www.joelcsaji.com/ai-dictionary/evals) (`evals`, Evaluation & Safety, Established): Tests that measure whether an AI system performs well on representative tasks or failure cases.
- [Golden Dataset](https://www.joelcsaji.com/ai-dictionary/golden-dataset) (`golden-dataset`, Evaluation & Safety, Established): A curated set of representative examples and expected outcomes used to evaluate an AI system.
- [Trace Grading](https://www.joelcsaji.com/ai-dictionary/trace-grading) (`trace-grading`, Evaluation & Safety, Rising): Scoring or labeling an agent's end-to-end trace to evaluate decisions, tool calls, and outcomes.
- [Orchestration Trace](https://www.joelcsaji.com/ai-dictionary/orchestration-trace) (`orchestration-trace`, Evaluation & Safety, Emerging): A trace that records coordination events in a multi-agent system, such as spawning sub-agents, delegating work, communicating, aggregating results, and stopping.
- [Judge Model](https://www.joelcsaji.com/ai-dictionary/judge-model) (`judge-model`, Evaluation & Safety, Rising): A model used to score, compare, or critique other model outputs.
- [Guardrails](https://www.joelcsaji.com/ai-dictionary/guardrails) (`guardrails`, Evaluation & Safety, Established): Rules, checks, filters, or workflows that keep AI behavior within allowed boundaries.
- [Tripwire](https://www.joelcsaji.com/ai-dictionary/tripwire) (`tripwire`, Evaluation & Safety, Rising): A guardrail signal that halts or blocks an agent run when input, output, or a tool action fails a policy check.
- [Ambient Authority](https://www.joelcsaji.com/ai-dictionary/ambient-authority) (`ambient-authority`, Evaluation & Safety, Rising): A security failure mode where an agent can use tools, credentials, or host permissions simply because they are present in its runtime.
- [Action-Time Authorization](https://www.joelcsaji.com/ai-dictionary/action-time-authorization) (`action-time-authorization`, Evaluation & Safety, Emerging): Checking whether a proposed agent action is authorized at the exact moment before the tool call, API request, or side effect executes.
- [Prompt Injection](https://www.joelcsaji.com/ai-dictionary/prompt-injection) (`prompt-injection`, Evaluation & Safety, Established): An attack or failure mode where untrusted content tries to override the system's intended instructions.
- [Tool Poisoning](https://www.joelcsaji.com/ai-dictionary/tool-poisoning) (`tool-poisoning`, Evaluation & Safety, Emerging): A threat where malicious or misleading tool metadata causes an agent to misuse tools.
- [Memory Poisoning](https://www.joelcsaji.com/ai-dictionary/memory-poisoning) (`memory-poisoning`, Evaluation & Safety, Emerging): A threat where false or malicious information is written into an agent's memory and affects future behavior.
- [Trace](https://www.joelcsaji.com/ai-dictionary/trace) (`trace`, Evaluation & Safety, Rising): A record of model calls, tool calls, inputs, outputs, timings, and decisions during an AI run.
- [Sandbox](https://www.joelcsaji.com/ai-dictionary/sandbox) (`sandbox`, Evaluation & Safety, Established): A restricted environment where agents can run tools or code without affecting production systems.
- [Hallucination](https://www.joelcsaji.com/ai-dictionary/hallucination) (`hallucination`, Evaluation & Safety, Established): A confident model output that is false, unsupported, or not grounded in the available evidence.
- [Frontier Model](https://www.joelcsaji.com/ai-dictionary/frontier-model) (`frontier-model`, Models & Training, Established): A model near the leading edge of current AI capability, often across reasoning, coding, multimodal, or agentic tasks.
- [Generative AI](https://www.joelcsaji.com/ai-dictionary/generative-ai) (`generative-ai`, Models & Training, Established): AI that creates new content such as text, images, audio, video, code, or structured data.
- [Large Language Model](https://www.joelcsaji.com/ai-dictionary/large-language-model) (`large-language-model`, Models & Training, Established): A model trained on large amounts of text and other data to understand and generate language.
- [Foundation Model](https://www.joelcsaji.com/ai-dictionary/foundation-model) (`foundation-model`, Models & Training, Established): A broadly trained model that can be adapted to many downstream tasks.
- [Multimodal Model](https://www.joelcsaji.com/ai-dictionary/multimodal-model) (`multimodal-model`, Models & Training, Established): A model that can process or generate multiple kinds of data, such as text, images, audio, video, or code.
- [Small Language Model](https://www.joelcsaji.com/ai-dictionary/small-language-model) (`small-language-model`, Models & Training, Rising): A smaller model optimized for lower cost, lower latency, local deployment, or narrow tasks.
- [Mixture of Experts](https://www.joelcsaji.com/ai-dictionary/mixture-of-experts) (`mixture-of-experts`, Models & Training, Established): A model architecture that routes each input through selected expert subnetworks instead of activating the whole model.
- [Fine-Tuning](https://www.joelcsaji.com/ai-dictionary/fine-tuning) (`fine-tuning`, Models & Training, Established): Training an existing model further on task-specific examples or domain data.
- [Distillation](https://www.joelcsaji.com/ai-dictionary/distillation) (`distillation`, Models & Training, Established): Training a smaller model to imitate the behavior or outputs of a larger model.
- [Synthetic Data](https://www.joelcsaji.com/ai-dictionary/synthetic-data) (`synthetic-data`, Models & Training, Established): Artificially generated data used for training, testing, evaluation, or simulation.
- [Quantization](https://www.joelcsaji.com/ai-dictionary/quantization) (`quantization`, Models & Training, Established): Reducing numerical precision in a model to make it smaller, faster, or cheaper to run.
- [AI Search](https://www.joelcsaji.com/ai-dictionary/ai-search) (`ai-search`, AI-Native Product, Rising): A search experience that uses generative AI to synthesize answers from retrieved sources instead of only listing links.
- [Voice Agent](https://www.joelcsaji.com/ai-dictionary/voice-agent) (`voice-agent`, AI-Native Product, Rising): An AI agent that listens and responds through speech, often while using tools or handing off tasks.
- [Vibe Coding](https://www.joelcsaji.com/ai-dictionary/vibe-coding) (`vibe-coding`, AI-Native Product, Rising): A casual term for building software by describing intent to AI coding tools and iterating on the generated result.
- [Coding Agent](https://www.joelcsaji.com/ai-dictionary/coding-agent) (`coding-agent`, AI-Native Product, Rising): An agent specialized for software tasks such as reading code, editing files, running tests, and debugging.
- [SWE-bench](https://www.joelcsaji.com/ai-dictionary/swe-bench) (`swe-bench`, AI-Native Product, Established): A benchmark family that evaluates AI systems on real software engineering issues.
- [Deep Research Agent](https://www.joelcsaji.com/ai-dictionary/deep-research-agent) (`deep-research-agent`, AI-Native Product, Rising): An agent workflow optimized for multi-step research, source gathering, synthesis, and citation-backed reports.
- [Agentic Commerce](https://www.joelcsaji.com/ai-dictionary/agentic-commerce) (`agentic-commerce`, AI-Native Product, Emerging): Commerce flows where AI agents discover products, compare options, negotiate, purchase, or support post-purchase actions.
- [Universal Commerce Protocol](https://www.joelcsaji.com/ai-dictionary/universal-commerce-protocol) (`universal-commerce-protocol`, AI-Native Product, Watch): An emerging open standard for agentic shopping flows across merchants, AI platforms, and commerce systems.
- [Agent Payments Protocol](https://www.joelcsaji.com/ai-dictionary/agent-payments-protocol) (`agent-payments-protocol`, Tools & Protocols, Watch): An emerging protocol for agent-led payments that focuses on consent, identity, auditability, and transaction authorization.
- [Mandate](https://www.joelcsaji.com/ai-dictionary/mandate) (`mandate`, Tools & Protocols, Watch): A signed instruction or authorization record that constrains what an agent is allowed to do in a transaction.
- [Agentic Web](https://www.joelcsaji.com/ai-dictionary/agentic-web) (`agentic-web`, AI-Native Product, Watch): The emerging idea that websites, apps, APIs, and permissions will be designed for agents as first-class users.
- [AI Coworker](https://www.joelcsaji.com/ai-dictionary/ai-coworker) (`ai-coworker`, AI-Native Product, Watch): A product framing where an AI system owns ongoing work streams rather than answering isolated prompts.
- [Agent Control Plane](https://www.joelcsaji.com/ai-dictionary/agent-control-plane) (`agent-control-plane`, AI-Native Product, Rising): A centralized layer for deploying, operating, monitoring, and governing many agents across an organization.
- [AgentOps](https://www.joelcsaji.com/ai-dictionary/agentops) (`agentops`, AI-Native Product, Emerging): Operational practices for deploying, monitoring, evaluating, and governing agentic systems.
- [Approval Gate](https://www.joelcsaji.com/ai-dictionary/approval-gate) (`approval-gate`, AI-Native Product, Rising): A product control that pauses an AI workflow until a person or policy approves the next action.
- [Jagged Frontier](https://www.joelcsaji.com/ai-dictionary/jagged-frontier) (`jagged-frontier`, AI-Native Product, Rising): The pattern where AI systems are excellent at some tasks while surprisingly weak at nearby tasks.
- [Extended Thinking](https://www.joelcsaji.com/ai-dictionary/extended-thinking) (`extended-thinking`, Reasoning & Inference, Rising): A model mode that allocates extra inference tokens to deliberate before answering, often with a configurable thinking budget.
- [Speculative Decoding](https://www.joelcsaji.com/ai-dictionary/speculative-decoding) (`speculative-decoding`, Reasoning & Inference, Rising): An inference technique where a smaller draft model proposes tokens that a larger model verifies in parallel to reduce latency.
- [Context Rot](https://www.joelcsaji.com/ai-dictionary/context-rot) (`context-rot`, Context & Memory, Emerging): The gradual loss of instruction-following quality as a conversation or agent trace fills with stale, noisy, or contradictory context.
- [Browser Use](https://www.joelcsaji.com/ai-dictionary/browser-use) (`browser-use`, Agents & Autonomy, Rising): An agent pattern that operates a real browser (clicking, typing, and reading pages) to complete tasks on the open web.
- [Prefix Caching](https://www.joelcsaji.com/ai-dictionary/prefix-caching) (`prefix-caching`, Reasoning & Inference, Rising): Reusing the computed key-value cache for a shared prompt prefix across requests so repeated system prompts and tools are cheaper and faster.
- [Lethal Trifecta](https://www.joelcsaji.com/ai-dictionary/lethal-trifecta) (`lethal-trifecta`, Evaluation & Safety, Rising): An agent combining private-data access, exposure to untrusted content, and external communication is practically undefendable against prompt injection.
- [Code Mode](https://www.joelcsaji.com/ai-dictionary/code-mode) (`code-mode`, Tools & Protocols, Rising): Instead of invoking tools one call at a time, the agent writes code that calls MCP servers exposed as code APIs, so only the definitions and results it actually needs enter context.
- [Tool Search](https://www.joelcsaji.com/ai-dictionary/tool-search) (`tool-search`, Tools & Protocols, Rising): A search layer over available tool definitions that loads only the few relevant ones on demand instead of front-loading every definition into the prompt.
- [Progressive Disclosure](https://www.joelcsaji.com/ai-dictionary/progressive-disclosure) (`progressive-disclosure`, Context & Memory, Rising): Loading agent context in stages - names and descriptions first, full bodies when relevant, linked files only as needed - so bundled context is effectively unbounded.
- [Spec-Driven Development](https://www.joelcsaji.com/ai-dictionary/spec-driven-development) (`spec-driven-development`, AI-Native Product, Rising): Defining requirements as an executable spec before coding - spec, plan, tasks, implement - with the AI coding agent implementing against it.
- [Agentic Engineering](https://www.joelcsaji.com/ai-dictionary/agentic-engineering) (`agentic-engineering`, AI-Native Product, Rising): Seasoned engineers accelerating their work with LLMs while staying accountable for the software they produce - the disciplined counterpart to vibe coding.
- [Diffusion Language Model](https://www.joelcsaji.com/ai-dictionary/diffusion-language-model) (`diffusion-language-model`, Models & Training, Emerging): A language model that generates entire blocks of tokens by iterative denoising rather than one token at a time.
- [World Model](https://www.joelcsaji.com/ai-dictionary/world-model) (`world-model`, Models & Training, Rising): AI systems that can simulate aspects of the world - Genie 3 generates interactive environments navigable in real time from a text prompt.
- [RL Environment](https://www.joelcsaji.com/ai-dictionary/rl-environment) (`rl-environment`, Models & Training, Rising): A packaged task world for reinforcement learning - the world, rules, and feedback loop of state, action, and reward a model trains inside.
- [Reward Hacking](https://www.joelcsaji.com/ai-dictionary/reward-hacking) (`reward-hacking`, Evaluation & Safety, Rising): An AI fooling its training or grading process into assigning high reward without completing the intended task - satisfying the letter of the objective, not its spirit.
- [x402](https://www.joelcsaji.com/ai-dictionary/x402) (`x402`, Tools & Protocols, Rising): An open standard for internet-native payments that revives HTTP 402 Payment Required: the server responds 402, the client pays instantly with stablecoins and retries.
- [GEO](https://www.joelcsaji.com/ai-dictionary/geo) (`geo`, AI-Native Product, Established): Optimizing content so AI-powered answers include you, describe you accurately, and recommend you for the right moments.
- [AI Slop](https://www.joelcsaji.com/ai-dictionary/ai-slop) (`ai-slop`, Evaluation & Safety, Established): Digital content of low quality produced usually in quantity by means of artificial intelligence - Merriam-Webster's 2025 Word of the Year.
- [Sleep-Time Compute](https://www.joelcsaji.com/ai-dictionary/sleep-time-compute) (`sleep-time-compute`, Context & Memory, Emerging): Letting models think during downtime - background agents use idle time to process information and rewrite their memory state between tasks.