Interaction#
What agents do: behavior, commands, communication, and the mechanisms that govern responses
accessibility tree snapshot#
Definition: structured, text-based representation of a web page’s roles, labels, descriptions, focus state, and interactive elements, pulled from the browser’s accessibility API rather than raw HTML or a screenshot
Purpose: gives an agent a compact, semantic view of a page’s interactive surface - far fewer tokens than raw DOM, more stable references than pixel coordinates - so the LLM can plan against addressable elements and the harness can enforce action-level safety checks on semantic labels; typically paired with vision as a fallback for canvas-based or non-DOM content the tree doesn’t expose
Example: a form field labeled “Total Weight (kg)” that requires 100+ lines of
nested HTML with ARIA attributes collapses to ref=38 heading "Total Weight (kg)..."
and ref=39 textbox "Total Weight (kg)..." focusable focused required, which the agent
references directly in a type tool call instead of parsing markup; Playwright MCP and
Chrome DevTools MCP both generate these snapshots instead of forwarding raw HTML
Related Terms: bulk action, harness, MCP server, permission and safety systems
Source: arXiv: “Building Browser Agents: Architecture, Security, and Practical Solutions” by Aram Vardanyan
AGENTS.md#
Definition: Markdown file placed at a repository’s root that gives coding agents project-specific build, test, and contribution instructions; best understood as “a README for agents”; jointly developed by OpenAI, Google, Sourcegraph, Cursor, and Factory, then donated to the Linux Foundation in December 2025 to keep governance neutral
Purpose: standardizes how a codebase communicates repo-specific behavior to any agent working in it - build commands, code style, testing instructions, security notes, commit and/or pull request rules - without vendor lock-in
Format: the spec has no required fields, but most repos include a project overview, build/test commands, and code style conventions
Related Terms: llms.txt, robots.txt,
rule, /SKILL,
spec, system prompt
Sources:
assistant message#
Definition: output generated by an agent or LLM during a conversational turn; paired with user messages to form the back-and-forth history the agent uses as context
Purpose: represents the agent’s side of an exchange and contributes to the accumulated conversation history that informs subsequent responses
Related Terms: turn, user message
Source: Dachary Carey: “An Agent is More Than Its Brain”
bulk action#
Definition: also known as action batching; grouping of multiple independent tool calls into a single dispatched request instead of executing them one at a time
Purpose: reduces network round-trips and LLM inference overhead for tasks that don’t require intermediate verification between steps, such as populating several fields in a form; contrasts with sequential execution, where the agent waits on each action’s result before issuing the next
Example: on a 28-field form, batching field fills and dropdown selections into one bulk action cut tool calls by 74%, 10 vs. 38, execution time by 57%, 104.5s vs. 245.1s, and tokens consumed by 41%, 154K vs. 260K, compared to issuing each field as a separate sequential action
Related Terms: accessibility tree snapshot, context window, harness
Source: arXiv: “Building Browser Agents: Architecture, Security, and Practical Solutions” by Aram Vardanyan
gate#
Definition: prompt condition that must be satisfied before a work sequence continues; provides objectively evaluable agentic checkpoints — thing happens → condition → proceed
Purpose: enforces deterministic progression that an LLM cannot loophole; contrasts with rules, which an LLM can interpret, bypass, or rationalize around, and with hooks, which are triggered by events in the harness
hook#
Definition: script or callback that runs automatically in response to a specific event in the agent’s environment; fires deterministically based on triggers, such as a file being edited, without going through the LLM’s interpretation loop
Purpose: enforces constraints reliably without relying on the agent to remember them; contrasts with gates, which block progression on a condition, and with rules, which an LLM applies at its own discretion
Related Terms: /SKILL, gate,
rule, slash command
Source: Dachary Carey: “An Agent is More Than Its Brain”
MCP server#
Definition: acronym for Model Context Protocol server; open-source standard for connecting AI applications to external systems; external server that exposes capabilities to an agent - tools, resources, and/or prompts
Purpose: lets agents interact with databases, APIs, cloud services, or custom systems the server is built to access; facilitates portable behavior across agent platforms because implementation lives in the server rather than the harness
Example: MCPs introduce security risk, Anthropic’s Git MCP server received three injection-related CVEs, common vulnerabilities and exposures, in 2025
Related Terms: accessibility tree snapshot, agent-friendliness, harness
Sources:
- ByteByteGo: “LLM Security Basics: The Full Threat Model”
- Dachary Carey: “An Agent is More Than Its Brain”
- LF Projects, LLC: “What is the Model Context Protocol (MCP)?”
observability#
Definition: formal framework within which agent evaluation is conducted; involves tracking, logging, and monitoring an agent’s behavior, decisions, and intermediate actions
Purpose: enables inspection of function calls, API responses, and retrieval steps to diagnose failures and understand what the agent actually did
Related Terms: permission and safety systems, self-reflection
planning#
Definition: agent capability to decompose complex goals into sequences of executable actions; fundamental building block for autonomous task execution
Purpose: involves reasoning about future states, choosing strategies, and organizing steps; ranges from basic linear plans to complex multi-step reasoning with contingencies
Related Terms: agent, memory, self-reflection
prompt#
Definition: input text or instructions given to an LLM to guide its response; quality and specificity of prompts significantly affect output quality
Purpose: is interpreted by the LLM (unlike direct commands) and is distinct from traditional search queries; related to the system prompt
Related Terms: /SKILL, system prompt,
user message
Promptions#
Definition: portmanteau of “prompt” and “options”; a lightweight middleware framework that dynamically generates ephemeral UI elements - radio buttons, checkboxes, toggles - to help users steer generative AI responses without manual prompt engineering
Purpose: lowers the barrier to effective prompting by replacing free-form text refinement with structured, parameterized choices; users select options and the AI response updates in real time
Example: a user asks an AI to explain a Python function; Promptions generates context-specific
controls like Experience Level: Beginner/Intermediate/Advanced,
Learning Objective: Training & Eval/Model Building/Data Processing, and
Response Format: Step-by-step/Summary/Detailed while the user adjusts these without retyping
the prompt
| Concept | Description |
|---|---|
| Dynamic Prompt Refinement Control | UI controls generated contextually from the user’s actual prompt input; adapts to the user’s request rather than offering generic options |
| Ephemeral UI | Interfaces generated on-the-fly by generative AI systems that exist just long enough to serve a specific purpose; not permanently stored |
| Prompt Middleware | Software layer between user and LLM that assists with prompt construction; the general category Promptions belongs to |
| Static Prompt Refinement Control | Fixed, generic set of controls - length, tone, format - offered regardless of prompt content; predictable, but not tailored |
Related Terms: harness, prompt, system prompt
Sources:
- GitHub: Promptions - Ephemeral UI for prompting
- Microsoft: “Tools for Thought - Better Thinking Through AI”
- Microsoft Research: “Dynamic Prompt Middleware: Contextual Prompt Refinement Controls for Comprehension Tasks” by Drosos et al.
rule#
Definition: prompt instruction an LLM interprets and applies at its own discretion; has an implicit opt-out path through which the LLM can rationalize skipping it
Purpose: guides behavior flexibly through LLM interpretation; contrasts with gates, which block progression until a condition is met, and with hooks, which fire deterministically from the harness regardless of LLM interpretation
Related Terms: gate, hook, permission and safety systems
self-reflection#
Definition: agent capability to evaluate its own reasoning, actions, and outputs; involves identifying errors, assessing performance, and adjusting strategy
Purpose: enables learning from mistakes and iterative improvement without external feedback; distinguishes more sophisticated agents from basic reactive systems
Related Terms: agent, memory, planning
/SKILL#
Definition: bundle of instructions and reference material that gives an agent just-in-time context for a specific domain or task; distinc from hooks or slash commands in that they work through LLM interpretation rather than deterministic execution, and subject to the same context window attention dynamics as other injected context
Purpose: extend an agent’s capabilities for specific domains or tasks by injecting relevant context at runtime, enabling more accurate and specialized behavior without modifying the core model or system configuration
Example: The Open Agent Skills Ecosystem is a community-driven registry
where developers publish, discover, and install /SKILL bundles for common agent tasks;
MongoDB Agent Skills provide pre-built skills for database
operations, allowing agents to interact with MongoDB collections and documents through structured instruction bundles
Related Terms: context window, hook, prompt, slash command, system prompt
Sources:
slash command#
Definition: direct command typed into a chat interface, such as /compact or /init;
triggers specific agent behavior without going through the LLM’s interpretation loop
Purpose: provides more predictable and consistent action execution than natural language prompts; contrasts with prompts, which the LLM interprets and may execute differently across runs
Related Terms: /SKILL, hook,
prompt
Source: Dachary Carey: “An Agent is More Than Its Brain”
system prompt#
Definition: set of instructions provided to the LLM by the platform before any user interaction begins; sits at the beginning of the context window, giving it strong positional attention weight
Purpose: typically not visible to the user, yet profoundly shapes the agent’s personality, default behaviors, and constraints
Related Terms: /SKILL, context window,
prompt, temperature
Source: Dachary Carey: “An Agent is More Than Its Brain”
turn#
Definition: single exchange in a conversation: one user message and one assistant message; the agent considers the full turn history when generating a response
Purpose: provides the unit structure of a conversation that the agent references for context
Related Terms: assistant message, context window, user message
Source: Dachary Carey: “An Agent is More Than Its Brain”
user message#
Definition: input sent by a human or automated system to an agent during a conversational turn; interpreted by the LLM rather than executed as a direct command
Purpose: receives strong positional attention as the most recent content in the context window, heavily influencing the agent’s response
Related Terms: assistant message, prompt, turn