Spec / 011 · terms

Agent spec dictionary

Short, consistent definitions for the vocabulary used across agents, guides, architectures, and pipelines. Terms are grouped by topic, cross-linked, and tied to catalog examples — not marketing gloss.

68 terms · Start with how agents work or OpenClaw vs Hermes Agent.

68 of 68 terms

Core agents

Agent: An AI agent is a system that uses a model plus tools and state to complete a task.
Agentic AI: Agentic AI refers to systems that take actions toward a goal, not only generate text.
Agentic Workflow: An agentic workflow is a sequence of observable stages an agent runs to finish a job.
Agent Architecture: Agent architecture is the arrangement of model, tools, memory, and control flow.
Single Agent: A single agent is one model-driven controller with tools and state for the whole task.
Tool Loop: A tool loop repeats plan → call tool → observe until a stop condition.
Stop Condition: A stop condition is the rule that ends an agent run: done, blocked, escalated, or timed out.
Planning: Planning is producing an inspectable next-step list or plan artifact.
Reflection: Reflection is a later pass that inspects an artifact and proposes a correction.
ReAct: ReAct is a pattern that interleaves explicit actions with observations.

Tools & retrieval

Tool Calling Also: function calling: Tool calling is when a model emits a structured request that a tool executes, then returns a result.
Function Calling: Function calling is a vendor name for tool calling with a JSON schema.
Observation: An observation is the structured result returned to the model after a tool runs.
Grounding: Grounding ties a claim to an observation: a quote, API field, or test result.
RAG: Retrieval-augmented generation fetches passages and then writes with them in context.
Retrieval: Retrieval is fetching candidate documents or records for a query.
Semantic Search: Semantic search ranks items by embedding similarity instead of exact keywords.
Vector Database: A vector database stores embeddings and returns similar items.
Embeddings: Embeddings are numeric vectors that represent text or other objects for similarity search.
Knowledge Base: A knowledge base is the corpus an agent may retrieve from: docs, tickets, or policies.
MCP Also: Model Context Protocol: Model Context Protocol (MCP) is an open standard for connecting agents to tools and context servers.
Agent Protocol: An agent protocol is a contract for how hosts and tools exchange structured actions.
Computer Use: Computer use is when an agent drives a GUI: click, type, scroll, and read the screen.
Structured Output: Structured output is a response constrained to a schema: JSON, enum, or grammar.

Memory & context

Memory: In agents, memory is stored information that can be retrieved later.
Short-Term Memory: Short-term memory is the working set for the current run: notes, tool results, and the active goal.
Long-Term Memory: Long-term memory is durable notes stored across sessions.
Memory Store: A memory store is the durable system that saves and retrieves agent notes.
Context: Context is the information placed in the model call for this step.
Context Engineering: Context engineering is choosing what information enters each model call.

Multi-agent

Handoff: A handoff is passing an artifact and responsibility from one agent to another.
Multi-Agent System: A multi-agent system is a designed workflow with more than one agent role.
Orchestration: Orchestration is control flow that assigns work, collects artifacts, and decides when to stop.
Orchestrator: An orchestrator is the role that routes tasks, validates handoffs, and stops the run.
Workflow Engine: A workflow engine runs declared stages, retries, and branches between agents or tools.
Supervisor Agent: A supervisor agent assigns work to workers and decides when to stop.
Worker Agent: A worker agent performs one specialist job and returns an artifact.
Router Agent: A router agent chooses which specialist or tool should handle a request.
Evaluator Agent: An evaluator agent scores or comments on an artifact against criteria.
Critic Agent: A critic agent is an evaluator that emphasizes defects and missing evidence.
Testing Agent: A testing agent exercises another system and reports pass, fail, or regressions.
A2A Also: agent-to-agent: Agent-to-agent (A2A) protocols define how separate agent hosts exchange tasks and artifacts.

Safety & ops

Guardrail: A guardrail is a constraint that blocks unsafe or out-of-policy actions.
Human-in-the-Loop: Human-in-the-loop means a person must confirm or take over before a side effect.
Agent Security: Agent security is protecting tools, credentials, and users from misuse by models or prompts.
Prompt Injection: Prompt injection is untrusted text that tries to override system instructions.
Sandbox: A sandbox is an isolated environment where agent code or commands run with limits.
Verification: Verification is an explicit check that an artifact meets criteria before it ships.
Agent Evaluation: Agent evaluation measures whether a system’s observable behavior meets a task spec.
Agent Observability: Observability is the ability to inspect inputs, tool calls, results, and outputs of a run.
Trace: A trace is the ordered log of steps in an agent run: inputs, tools, observations, outputs.
Eval Harness: An eval harness runs many agent or prompt variants against fixtures and assertions.
Checkpoint: A checkpoint is a saved state an agent run can resume from after failure or pause.
DevOps: DevOps for agents is how you deploy, monitor, and roll back agent systems in production.

Prompt workshop

Prompt Engineering: Prompt engineering is designing instructions, examples, and output contracts — and, in the workshop, compiling, testing, constraining, or optimizing them.
Compiled Prompt: A compiled prompt is the artifact an optimizer emits after searching instructions and demonstrations against a metric.
DSPy: DSPy is a framework for compiling language-model programs from signatures, modules, and a metric.
Prompt Evaluation: Prompt evaluation is running prompt variants against assertions and comparing the matrix to a baseline.
promptfoo: promptfoo is an eval harness for prompts and models: configs, assertions, diffs, and CI.
Constrained Decoding: Constrained decoding lets the sampler emit only tokens that match a grammar, enum, or regex.
Guidance: Guidance is a generation programming model that mixes fixed text with constrained holes.
Textual Gradient: A textual gradient is natural-language advice about how to change a prompt or other text parameter to reduce a language loss.
TextGrad: TextGrad optimizes textual variables by a forward generation, a language loss, a textual gradient, and an update.
GEPA: GEPA is reflective prompt evolution with Pareto selection: seed, execute, reflect, mutate, keep the non-dominated archive.

Coding agents

Vibe Coding: Vibe coding is directing software work in natural language while agents and tools change the repository.
AI Coding Agent: An AI coding agent edits software with repository and test tools in a loop.
Agentic Coding: Agentic coding is software development performed as an agent workflow: plan, edit, observe, repair.
Code Review: Code review is an evaluator pass that inspects a diff against style, safety, and requirements.