AI Agent Glossary
Plain-English definitions for agentic AI — the shared vocabulary, the enterprise vocabulary, and the terms unique to Nagent.
Core Agentic Vocabulary
The shared language of agentic AI — concepts you'll encounter in any modern agent system.
A
- Agent
- An autonomous software entity powered by AI that can perceive inputs, reason about goals, use tools, and take actions to complete tasks.
- Agent Architecture
- The structural design of an AI agent, including its memory, planning system, tools, reasoning engine, and execution loop.
- Agent Builder
- A platform or framework used to create, configure, and deploy AI agents without building everything from scratch.
- Agent Orchestration
- The coordination of multiple specialised AI agents working together on a shared goal, with routing, sequencing, and error-handling logic managing the flow.
- Agent Runtime
- The execution environment where AI agents operate, manage tasks, interact with tools, and maintain state.
- Agent Stack
- The complete set of technologies powering an AI agent — including models, memory, orchestration, APIs, retrieval systems, and infrastructure.
- Agent Swarm
- A large collection of specialised AI agents dynamically collaborating on parallel subtasks toward a shared objective.
- Agentic AI
- AI systems capable of autonomous, multi-step action — perceiving context, planning, and executing tasks without human intervention at each step.
- AI Copilot
- An AI assistant embedded within software products to augment human workflows through suggestions, automation, and contextual assistance.
- AI Governance
- Policies, controls, and monitoring systems used to ensure AI operates safely, ethically, securely, and within organisational boundaries.
- AI Native
- Products or workflows fundamentally designed around AI capabilities rather than retrofitting AI into legacy systems.
- Autonomous Execution
- The ability of an AI agent to carry out a defined task from start to finish without requiring human approval at each step.
B
- Benchmarking
- The process of evaluating AI model or agent performance against standard datasets, tasks, or quality metrics.
- Browser Agent
- An AI agent capable of interacting with websites like a human — clicking buttons, filling forms, scraping information, and navigating interfaces.
C
- Chain-of-Thought (CoT)
- A reasoning technique where the model explicitly generates intermediate reasoning steps before arriving at a final answer.
- Cognitive Architecture
- The high-level reasoning framework governing how an AI agent plans, learns, remembers, and makes decisions.
- Context Engineering
- The practice of designing what information, memory, instructions, and retrieved data are supplied to an agent during execution.
- Context Window
- The maximum amount of text or information an LLM can process at one time.
- Continuous Learning
- A system capability where agents improve over time through feedback, usage patterns, evaluations, or new training data.
- CrewAI
- An open-source framework for coordinating collaborative multi-agent workflows.
D
- Decision Engine
- The logic layer responsible for determining which actions, tools, or agents should be activated during execution.
- Deep Research Agent
- An AI agent designed for long-form web research, synthesis, source comparison, and insight generation across multiple data sources.
- Deterministic Workflow
- A workflow where outputs follow predictable rule-based logic rather than probabilistic reasoning.
- Diffusion Model
- A type of generative AI model commonly used for image, video, and media generation.
E
- Embedding
- A numerical representation of text, images, or data that enables semantic search and similarity matching.
- Evaluation (Evals)
- Automated or human-defined tests used to measure the quality, safety, accuracy, or usefulness of AI outputs.
- Execution Loop
- The recurring cycle where an agent observes, reasons, acts, evaluates, and repeats until a goal is completed.
F
- Feedback Loop
- A mechanism where agent outputs are evaluated and used to improve future performance — either through human signals or automated evaluation.
- Few-Shot Prompting
- A prompting method where examples are provided within the prompt to guide model behaviour.
- Fine-tuning
- The process of updating a language model's weights using domain-specific training data to improve performance on targeted tasks.
- Foundation Model
- A large pre-trained AI model that serves as the base layer for many downstream applications and specialised systems.
- Function Calling
- A capability in modern LLMs that allows the model to invoke external tools, APIs, and services in a structured, predictable way.
G
- Grounding
- The process of connecting model outputs to real-world data sources, APIs, or verified knowledge to improve factual accuracy.
- Guardrails
- Safety constraints and policy mechanisms that prevent AI agents from producing harmful, insecure, or non-compliant outputs.
H
- Hallucination
- When a language model generates content that is factually incorrect or fabricated, presented with apparent confidence.
- Human-in-the-Loop (HITL)
- A workflow design pattern where an AI agent pauses at defined checkpoints to wait for human review or approval before proceeding.
- Hybrid AI System
- An AI architecture combining symbolic systems, deterministic logic, and probabilistic language models.
- Hyperparameter
- A configurable setting used during AI model training or inference that influences behaviour and performance.
I
- Inference
- The process of generating outputs from a trained AI model during runtime.
- Intent Detection
- The capability of an AI system to identify the user's underlying objective or goal from natural language input.
J
- JSON Mode
- A structured output mode where an LLM generates responses in machine-readable JSON format.
K
- Knowledge Base
- A structured repository of domain-specific information that agents can retrieve and reference when executing tasks.
L
- LangChain
- An open-source framework for building applications powered by language models, widely used for constructing agent pipelines.
- Latency
- The time taken for an AI system or agent to produce a response after receiving an input.
- Latent Space
- A compressed mathematical representation where AI models encode patterns, relationships, and semantic meaning.
- LLM (Large Language Model)
- A deep learning model trained on large text corpora that can generate, classify, summarise, and transform natural language.
- Long-Term Memory
- Persistent memory retained across multiple sessions, workflows, or users.
M
- MCP (Model Context Protocol)
- A standardised protocol enabling AI agents to securely connect with external tools, systems, APIs, and enterprise data sources.
- Memory (Agent Memory)
- The ability of an agent to retain and recall information from prior interactions — enabling personalisation, consistency, and context-awareness across sessions.
- Model Alignment
- Techniques used to ensure AI systems behave according to human goals, values, and safety expectations.
- Model Distillation
- The process of transferring capabilities from a larger model into a smaller, faster, or cheaper model.
- Model Routing
- The process of dynamically selecting the most appropriate AI model for a given task based on cost, speed, or capability.
- Multi-Agent System
- An architecture where multiple specialised AI agents collaborate on a task, each handling the part of the problem it is best suited for.
- Multi-Step Reasoning
- The capacity of an AI agent to decompose a complex goal into sub-tasks, reason through them in sequence, and synthesise a final output.
- Multimodal AI
- AI systems capable of understanding and generating across multiple input types — text, images, audio, video, and code.
N
- Neural Network
- A machine learning architecture inspired by the human brain, forming the foundation of modern deep learning systems.
- Neuro-Symbolic AI
- An approach combining neural networks with symbolic reasoning and logic systems.
O
- Observability
- The ability to monitor, inspect, debug, and understand how AI agents make decisions and execute workflows.
- On-Premise Deployment
- Hosting AI infrastructure within an organisation's own servers and private environments instead of public cloud platforms.
P
- Persona Conditioning
- The practice of shaping an AI agent's communication style, tone, expertise, or identity through system instructions.
- Planning Engine
- The component responsible for decomposing goals into executable steps and sequencing actions intelligently.
- Programmatic SEO
- The technique of generating large volumes of SEO-optimised pages from structured data, targeting long-tail search queries at scale.
- Prompt Chaining
- A technique where outputs from one prompt become inputs to subsequent prompts in a workflow.
- Prompt Engineering
- The practice of crafting and optimising the instructions given to an LLM to elicit more accurate, relevant, or structured outputs.
- Prompt Injection
- A security vulnerability where malicious instructions manipulate an AI system's behaviour through crafted inputs.
Q
- Quantisation
- A model optimisation technique that reduces model size and computational cost by lowering numerical precision.
R
- RAG (Retrieval-Augmented Generation)
- An architecture that combines a language model with a retrieval system — the model queries a knowledge base to ground its responses in relevant, up-to-date information.
- ReAct (Reasoning + Acting)
- A prompting pattern where an LLM alternates between reasoning steps and tool-use actions, mimicking how a human might tackle a problem.
- Reasoning Model
- An AI model optimised for logical thinking, planning, analysis, and multi-step problem solving.
- Recursive Self-Improvement
- A capability where AI systems iteratively refine their own outputs, prompts, or strategies.
- Reinforcement Learning
- A machine learning paradigm where models learn by receiving rewards or penalties based on actions.
- Reinforcement Learning from Human Feedback (RLHF)
- A model training technique where human preferences guide model optimisation.
- Retrieval Pipeline
- The system responsible for fetching relevant documents, embeddings, or knowledge before generation occurs.
S
- Sandbox Environment
- An isolated execution environment used for safely testing AI-generated code or workflows.
- Semantic Search
- Search based on meaning and contextual similarity rather than exact keyword matching.
- Session Memory
- Temporary conversational memory retained during an active interaction session.
- Small Language Model (SLM)
- A compact AI language model optimised for efficiency, speed, or on-device deployment.
- Sovereign AI
- An approach to AI deployment where an organisation retains full control over their AI infrastructure, data, and models — typically via private cloud or on-premise deployment.
- Synthetic Data
- Artificially generated data used for model training, testing, or simulation.
- System Prompt
- Instructions passed to an LLM before the user message that define the agent's role, constraints, tone, and behaviour.
T
- Task Decomposition
- The process of breaking complex objectives into smaller executable subtasks.
- Token
- The basic unit of text processed by a language model — roughly equivalent to a word or word-part. Context limits and costs are measured in tokens.
- Tool Registry
- A catalogue of APIs, services, databases, and utilities available for agents to invoke.
- Tool Use (Function Calling)
- The ability of an LLM to invoke external functions — APIs, databases, calculators, code runners — and incorporate the results into its response.
- Toolformer
- A model architecture designed to autonomously learn how and when to use tools.
- Transformer Architecture
- The neural network architecture underlying modern LLMs like GPT, Claude, and Gemini.
U
- Unstructured Data
- Information without predefined schema — including PDFs, emails, images, audio, video, and free-form text.
V
- Vector Database
- A specialised database optimised for storing and retrieving embeddings used in semantic search and RAG systems.
- Vector Search
- Similarity-based retrieval using embedding vectors rather than exact text matching.
- Voice Agent
- An AI agent capable of understanding and responding through spoken language.
W
- Web Agent
- An AI agent capable of navigating websites, extracting information, and executing browser-based workflows autonomously.
- Workflow Automation
- The use of software (including AI agents) to execute a repeatable sequence of business tasks with minimal human intervention.
- Workflow Engine
- The orchestration system managing execution order, dependencies, retries, and state transitions within AI workflows.
X
- XAI (Explainable AI)
- Methods and systems designed to make AI decisions understandable and interpretable to humans.
Z
- Zero-Shot Prompting
- A prompting technique where a model performs a task without prior examples in the prompt.
Advanced & Enterprise
Operational and architectural terms that show up when agents move into production.
A
- AEO (Answer Engine Optimization)
- Optimising content so AI answer engines (ChatGPT, Perplexity, Claude, Google AI Overviews) cite or summarise it correctly. The successor discipline to SEO when the consumer is a model, not a search results page.
- Agent Evaluation Pipeline
- An automated test harness that scores agent outputs against ground-truth tasks, brand-voice rules, and business outcomes — re-run on every prompt or model change so regressions surface before production.
- Agent Observability
- The instrumentation and tooling that let operators see what an agent did, why, and at what cost — including prompts, tool calls, latency, and final outputs.
- Agent Reliability
- A measure of how consistently an agent produces correct, on-policy outputs across repeated runs and edge cases. Tracked over time to detect drift and reward stability.
- Agent Telemetry
- The structured event stream emitted by an agent during execution — token spend, tool invocations, errors, and stage transitions — that powers dashboards and alerting.
- Autonomous Planning
- An agent's ability to choose its own sequence of steps to reach a goal, rather than following a pre-defined script. Combines reasoning, tool selection, and self-correction.
C
- Cognitive Loop
- The repeating cycle of perceive → reason → act → reflect that an agent runs until its goal is met or a stop condition fires.
D
- Dynamic Tool Routing
- The runtime decision of which tool, API, or sub-agent to invoke for a given step — based on the input, current context, and prior outcomes rather than a hard-coded path.
E
- Enterprise RAG
- A retrieval-augmented setup hardened for organisational use — with permissions, source attribution, freshness controls, and multi-source fusion across internal documents, databases, and SaaS systems.
- Execution Graph
- A directed graph of the actual steps an agent took — nodes are tool calls or reasoning checkpoints, edges record causality. Used for debugging, replay, and audit.
G
- Goal-Oriented Planning
- A planning approach where the agent works backward from a defined outcome, decomposing it into the actions and sub-goals required to reach it.
H
- Human Preference Optimisation
- A training method that uses pairwise human judgements ("A is better than B") to fine-tune model behaviour toward preferred outputs. Underlies modern instruction-following models.
K
- Knowledge Graph
- A structured network of entities and the relationships between them, used to give agents grounded, queryable context that complements free-text retrieval.
M
- Memory Compression
- Techniques for shrinking long agent histories into shorter summaries or embeddings so the relevant context fits within model token limits without losing essential signal.
- Model Arbitration
- The runtime decision of which underlying model (e.g. Sonnet vs. Haiku, open vs. closed) to use for a given step — balancing cost, latency, and capability.
- Multi-Modal Reasoning
- An agent's ability to combine inputs across text, images, audio, and structured data within a single reasoning chain — for example, reading a chart and explaining its implications.
P
- Policy Engine
- The layer that enforces what an agent is allowed to do — which tools it may call, which data it may read, which actions need human approval — independent of the model itself.
R
- Retrieval Cache
- A short-lived store of recently fetched documents and embeddings that lets repeat queries skip the expensive retrieval round-trip and stay cheap under load.
- Runtime Context Injection
- The mechanism that assembles a fresh context window per turn — pulling user history, retrieved docs, system prompt, and tool schemas into the model call dynamically.
S
- Self-Healing Workflow
- A workflow design where failed steps trigger automatic retries, fallbacks, or escalation paths so the overall job completes despite transient errors.
- Sovereign AI
- Deployment architecture where AI runs entirely inside the customer's environment — bring-your-own-model, on-prem or single-tenant cloud, data never leaves the perimeter. Required for regulated industries and the default for Nagent Enterprise customers.
- Stateful Agent
- An agent that retains and reuses information across turns or sessions — as opposed to a stateless agent that treats every request as fresh with no memory.
V
- VisibilityOS
- A discipline + tooling stack for monitoring how AI answer engines describe your brand, products, and competitors over time. Includes citation tracking, share-of-answer benchmarks, and continuous content tuning. The category Nagent's AEO module addresses.
T
- Tool Arbitration
- The decision logic that resolves conflicts when more than one tool could satisfy a step — picking the cheapest, fastest, or most authoritative option.
- Tool-Augmented Reasoning
- A pattern where an LLM extends its reasoning by calling external tools mid-thought — such as a calculator, search index, or code runner — and folds the result back into its answer.
W
- Workflow Graph
- The declarative blueprint of a multi-step process — nodes are tasks or agents, edges define dependencies and data flow. The runtime executes the graph.
- World Model
- An agent's internal representation of how its environment behaves — what tools exist, what users want, what state the system is in — used to predict the consequences of its actions.
Nagent-Proprietary
Vocabulary unique to the Nagent platform. These describe how our agents are built, governed, and deployed.
A
- Action Registry
- The typed catalogue of every action a Nagent agent can perform. Helix only suggests tools that exist here; BuildCraft compiles against the same registry; governance gates per-action overrides against it.
- Action Spine
- The orchestration backbone responsible for routing actions, tools, and approvals across agents.
- Agent DNA
- The unique behavioural profile defining how an agent thinks, communicates, reasons, and acts.
- Agent Genome
- The complete configuration blueprint defining an agent's personality, memory, tools, permissions, and reasoning behaviour.
- Agent Smriti
- Nagent's persistent memory layer that stores contextual knowledge, user preferences, and interaction history across sessions and agents. See also: Smriti.
- AgentSpec
- The shared schema every Nagent agent conforms to — identity, behaviour (system prompt + model), capabilities (tools + skills + events), governance (autonomy level, risk level, action overrides, guardrails), triggers, and budget. Same shape whether the agent was authored in Helix, BuildCraft, or pulled from the marketplace.
- Agentic AI Lab
- Nagent's in-house team that designs, builds, and deploys a fully custom agent system for the customer end-to-end — used for enterprise, regulated, or large-scale deployments.
- Approval Engine
- Multi-party approval workflow built into the platform. Flagged actions (e.g. any outbound to a C-level contact, any spend above a threshold) queue for human review before firing. Captures every decision in the audit trail.
- ARIA
- Function specialist on the Nagent entry-agent team — Agentic Research & Intelligence Agent. Handles market research, competitive intel, trend analysis, and continuous insight delivery. Reports to MIRA.
- AuditLog
- The append-only event collection that records every write operation across the platform. Immutable hooks, 365-day retention. The evidentiary substrate compliance teams need to certify the platform.
- Autonomy Ladder (L0 → L4)
- Five-step ladder controlling how much an agent can do on its own. L0 = always ask. L1 = act on low-risk, ask on the rest. L2 = act on most, request approval on flagged actions. L3 = act autonomously, log everything. L4 = full autonomy with audit-trail review only. Operators can demote or promote any agent in one click.
B
- Banking Strategist
- Industry strategist on the Nagent entry-agent team. Cross-functional within Banking & Financial Services — KYC/AML, risk scoring, complaint resolution, advisory bots, cross-sell. Reports to MIRA.
- Brand Lock AI
- Nagent's free public tool. Paste a URL + logo + reference materials, get a complete brand book (12 cognition engines + radar diagram + printable PDF). Doubles as a lead-magnet and demonstrates the brand substrate that paid Brand Workspaces use internally.
- Brand Workspace
- A fully isolated tenant — own brand book, governance, model preferences, audit trail. Built for agencies, multi-brand enterprises, and partner ecosystems. Operators switch workspaces with one click; permissions are scoped per-workspace.
- BuildCraft
- Nagent's pro-code IDE for agents — TypeScript / Python SDKs, full Git workflow, custom tools, deep external integrations. Use when Helix's no-code surface isn't enough or you need IP-level customisation.
C
- CEVA
- Function specialist on the Nagent entry-agent team — Customer Experience & Value Agent. Handles support automation, retention, onboarding, NPS, churn, and win-back. Reports to MIRA.
- Cockpit
- The real-time fleet view inside the Topology surface. Shows every agent, every action, every handoff, every queued approval as it happens. Operators can pause, demote, isolate, or override any agent without losing the run state.
- Cognitive Mesh
- Nagent's interconnected intelligence layer enabling agents to collaborate and share contextual awareness.
- Cognitive Router
- A system that dynamically selects the best model, tool, or agent for a task in real time.
- Composio Integration
- The broad-coverage integration layer Nagent uses to reach 800+ third-party tools (CRMs, email providers, chat, data warehouses, marketing automation). Composio handles the connector plumbing; Nagent's agents handle the orchestration.
- Content Studio
- Nagent's concierge content team — ad creative, video, brand-grade copy at agentic speed. Use when you need outcomes (campaigns / launches) rather than infrastructure.
- Cost Cap
- A guardrail type that bounds per-agent or per-action spend. Daily ceilings, per-run token caps, per-debate dollar caps. When breached the action queues for approval or fails-soft rather than continuing to spend.
- CREA
- Function specialist on the Nagent entry-agent team — Creative Content Execution Agent. Handles long-form articles, social posts, email sequences, video scripts, multilingual content. Reports to MIRA.
D
- Deal Room
- A shareable per-lead micro-site. The operator creates one with a click; the prospect gets a private URL they can visit without auth. Includes the proposal, relevant marketplace agents, case studies, and a booking link. Every visit + click is tracked back to the Lead timeline.
- Decisions Ledger
- The append-only record of every AI-suggested action (recommend stack, suggest next email, propose autonomy change) logged against the entity it affected and the outcome it produced. Operators can replay the chain to understand why a decision was made and what it led to.
- Multi-Agent Debate
- A moderated argument between specialist agents (e.g. Decision Engine vs Cost-Cap vs Brand-Voice) before high-stakes actions fire. Round-robin cadence, hybrid stop condition (consensus or N rounds), full transcript logged, spend-capped. Unique to Nagent — most platforms ship a single-agent execution model.
E
- EmailBuilder
- The variant generator + ranker for email templates. Given a context (lead, persona, intent), produces 3-5 candidate drafts and ranks them by predicted reply rate. The operator picks one; the choice flows back through Karmic so promotion is automatic over time.
- Entry Agent
- A customer-facing Nagent agent that visitors talk to on the public website. The entry-agent team is one super coordinator (MIRA), eight function specialists (MOXA, SERA, CEVA, ARIA, OPRA, CREA, TARA, PERA), and eight industry strategists (Banking, FMCG, Insurance, Pharma, SaaS, Real Estate, Fashion, Media).
- ETIA
- Email Template Intelligence Agent. Indexes every template in the library against persona / funnel-stage / tone / outcome history. When an operator (or another agent) needs a template for a given context, ETIA matches, resolves, and adapts the best candidate — and produces an explanation for the pick.
- Execution Context
- Snapshots of the context-evaluation pass the runtime took before invoking a model. Records which Smriti documents were pulled, which recent actions were considered, what state of related entities was assembled. Retained 30 days for debugging.
- Execution Memory
- A persistent record of previous actions, outcomes, and workflow states used for adaptive optimisation.
F
- FMCG Strategist
- Industry strategist on the Nagent entry-agent team. Cross-functional within FMCG / CPG — demand sensing, trade-promotion ROI, route-to-market, shopper insights. Reports to MIRA.
- Fashion Strategist
- Industry strategist on the Nagent entry-agent team. Cross-functional within Fashion & Apparel — trend forecasting, SKU lifecycle, omnichannel personalisation, returns analytics. Reports to MIRA.
G
- Governance Console
- The admin surface where operators control per-agent autonomy levels, per-action overrides, kill switches, approval workflows, and guardrail enforcement. The trust layer made operable.
- Guardrail
- A per-agent or per-action constraint. Seven kinds: cost-cap, content-restriction, approval-required, audience-restriction, tool-restriction, time-window, custom. Configurable per agent without touching the model layer.
H
- Handoff
- A formal cross-agent transition in the Nagent entry-agent team. The active agent emits a [handoff:slug|reason] marker; the audit log records the hop; the next turn comes from the receiving agent. Loop detection, disabled-target gate, session cap, and PII redaction all apply server-side.
- Helix
- Nagent's chat-first agent builder. Describe an agent in plain English — Helix synthesises the complete AgentSpec (name, system prompt, autonomy level, action permissions, triggers, skills) and ships it into the workbench. A deterministic auto-fix pass catches and repairs known-bad values before persist so the spec is always shippable.
I
- Industry Strategist
- Cross-functional entry agent anchored to a specific vertical (Banking, FMCG, Insurance, Pharma, SaaS, Real Estate, Fashion, Media). Speaks marketing / sales / ops / CX through the industry's lens.
- Insurance Strategist
- Industry strategist on the Nagent entry-agent team. Cross-functional within Insurance — claims triage, underwriting assist, FNOL automation, fraud detection. Reports to MIRA.
- Intent Graph
- A dynamic map of user goals, tasks, and contextual dependencies used for intelligent planning.
K
- KARMIC Feedback Loop
- Nagent's outcome-attribution + feedback loop. Every agent action is measured against business outcomes (replies, meetings booked, deals closed). Drafts that win get promoted. Agents that drift get demoted automatically. Operators see a trust score per agent and can intervene; they don't have to babysit the loop.
- Kill Switch
- One operator click pauses an agent, an agent class, or the entire fleet mid-flight. Pending actions queue; in-flight actions complete without triggering downstream. The single most-asked-for feature on every enterprise security review.
L
- Lead Sequence
- A multi-touch email cadence engine — cron-driven sends, follow-up + breakup logic, frequency caps, AB-variant promotion, reply handlers. Sequences are first-class entities with their own Smriti, Karmic feedback, and governance bounds.
- Live Ops
- The real-time operating-mode of the Topology cockpit. Shows agents working right now: what they're executing, what they just decided, where the next handoff is queued, which approvals are pending.
- Lookalike Profile
- An auto-synthesised ideal-prospect profile derived from the operator's declared ICP (or from closed-won customers when available). Used by NORA and the prospecting pipeline to score and rank discovered contacts.
M
- Media Strategist
- Industry strategist on the Nagent entry-agent team. Cross-functional within Media & Publishing — content recommendation, ad-yield optimisation, paywall lift, subscriber retention. Reports to MIRA.
- Memory Fabric
- Nagent's unified memory architecture combining short-term, long-term, and organisational memory.
- MIRA
- Super coordinator on the Nagent entry-agent team — Market Intelligence for Revenue Acceleration. First point of contact for every visitor on the public website. Reads platform context, routes the visitor to the right function specialist or industry strategist via [handoff] markers, and monitors the conversation so nothing falls through the cracks. Does NOT do function-specific work herself.
- MOXA
- Function specialist on the Nagent entry-agent team — Marketing Operations & Experience Agent. Handles campaigns, content, SEO, email, social, brand, demand-gen, ABM, analytics. Reports to MIRA.
N
- NORA
- Nagent's production outbound prospecting agent. 22 tools across DNA / discovery / research / drafting / sending / replies / cadence / pipeline. ICP-driven, cost-capped, outcome-attributed. Operates end-to-end without operator intervention at L4. Sub-agent of SERA on the entry-agent team.
O
- OPRA
- Function specialist on the Nagent entry-agent team — Operations & Process Automation Agent. Handles workflow automation, compliance, procurement, reconciliation, vendor management. Reports to MIRA.
P
- PERA
- Function specialist on the Nagent entry-agent team — Product & Engineering Research Agent. Handles user research synthesis, PRDs, roadmap prioritisation, competitive product intel. Reports to MIRA.
- Pharma Strategist
- Industry strategist on the Nagent entry-agent team. Cross-functional within Pharma & Healthcare — MLR review, MSL field intelligence, HCP engagement, adverse-event monitoring. Reports to MIRA.
- Pre-Meeting Prep Card
- An auto-generated five-minute briefing the operator gets before every Cal.id booking on a lead. Who you're meeting, what's happened in the relationship, the open questions, the likely objections, the suggested next step. Drops into the calendar invite or email.
R
- Real Estate Strategist
- Industry strategist on the Nagent entry-agent team. Cross-functional within Real Estate — site shortlisting, tenant lifecycle, property-doc Q&A, lead routing. Reports to MIRA.
S
- SaaS Strategist
- Industry strategist on the Nagent entry-agent team. Cross-functional within B2B SaaS — PLG onboarding, churn prediction, expansion plays, developer support. Reports to MIRA.
- Sales Assistant
- The inbound counterpart to NORA. Multi-turn conversation with leads via chat and email, with structured capture (intent, budget, timeline, decision process) that auto-advances the lead through the pipeline. Includes voice input, anti-hallucination validators, and a role-lens prompt extension that frames every output for the recipient's role.
- SERA
- Function specialist on the Nagent entry-agent team — Sales Execution & Research Agent. Handles prospecting, outreach, qualification, CRM, pipeline, forecasting, proposals. Reports to MIRA. NORA is her sub-agent.
- Skill
- A reusable markdown prompt module attached to one or more agents. Defines when to use it (trigger description) and how to apply it (the prompt body), with optional allowedTools[] to restrict which actions it can compose. Stored as .md-style records and editable in the admin Skills catalogue.
- Skill Capsule
- A reusable packaged capability combining prompts, tools, memory, and workflows into deployable agent skills.
- Smriti
- The persistent per-entity memory fabric. Every agent, lead, account, deal, operator, and conversation has its own structured .md-style document that grows as the entity accumulates history. Cross-thread, cross-agent, with background synthesis passes that compress history into higher-level patterns. The layer that lets Nagent agents remember.
- Stack (Agent Stack)
- A multi-agent composition. Visual orchestration graph wired in React Flow — output of one agent feeds the next, with conditional branches, parallel fan-out, and handoffs. Stacks are first-class entities with their own Smriti, Karmic feedback, and governance bounds.
T
- TARA
- Function specialist on the Nagent entry-agent team — Talent & HR Automation Agent. Handles recruiting, onboarding, engagement, performance reviews, L&D. Reports to MIRA.
- Topology
- The visual fleet view. Every agent in the deployment shown in a live graph with their current state, recent actions, and pending handoffs. Hover or click for the per-agent action log; pause, demote, isolate, or override any agent without losing the run.
- Trigger
- How an agent fires. Three kinds: event (canonical event published from anywhere in the platform), cron (standard 5-field cron expression), and manual (operator invocation). The event registry defines canonical names so triggers can't silently drift.
- Trust Layer
- Nagent's governance, permissions, auditability, and safety framework for enterprise AI deployment.
- Trust Score
- A per-agent quality score Karmic surfaces in the Governance Console. Composed of subscores for outcome attribution, action quality, hallucination rate, and operator-override frequency. Agents whose trust score drops get auto-demoted on the autonomy ladder until they earn the level back.