--- title: The vocabulary of agent email, defined metaTitle: Glossary of AI agent email terms description: SPF, DKIM, DMARC, MCP, threading, cold sends, suppression, idempotency. Every term that appears when you put an agent on an inbox, with what it means and why it matters. date: 2026-08-06 author: MCPmailer tags: Guides --- Agent email borrows vocabulary from two fields that developed separately: mail infrastructure, which is forty years old and full of acronyms, and agent tooling, which is new and full of jargon. Here is both, defined plainly, with a pointer to the piece that goes deeper. ![Terms from two fields meeting in one system](/blog/email-agent-glossary/hero.webp "Two vocabularies, one system, and a lot of acronyms.") ## Authentication and deliverability **SPF.** A DNS record listing which servers may send for a domain. Checked against the envelope sender, not the visible From address, which is why it alone proves little. **DKIM.** A cryptographic signature added by the sending domain, verified against a public key in DNS. Proves the message was signed by that domain and not modified in transit. **DMARC.** A policy saying what to do when SPF or DKIM does not **align** with the visible From domain. The one that actually matters, covered in [SPF, DKIM, and DMARC for an agent subdomain](/blog/dmarc-record-generator-agents). **Alignment.** Whether the domain that passed SPF or DKIM matches the From domain. Both can pass while DMARC fails, which is the single most common confusion in setup. **Reputation.** What mailbox providers think of your sending domain, based on history, volume shape, and how recipients react. Falls fast, recovers slowly, per [getting a damaged sending domain back](/blog/sending-reputation-recovery). **Warmup.** Growing volume gradually on a new domain so no history is not treated as suspicion, per [a warmup plan](/blog/email-warmup-plan-agents). **Hard bounce.** The address does not exist. Never send there again. **Soft bounce.** Temporary: full mailbox, server busy. Retry, then give up. **Complaint.** A recipient marked the message as spam. Measured per provider, and a rate rather than an incident is what matters. **Suppression.** A list of addresses that must not be contacted, enforced at the infrastructure layer rather than by an agent remembering, per [bounces, complaints, and suppression](/blog/bounces-suppression-agent-email). **Placement.** Whether delivered mail landed in the inbox, promotions, or spam. Different from delivery, and only visible through seed accounts, per [knowing where your agent's mail actually lands](/blog/deliverability-monitoring-seed-testing). ## Mail mechanics **Message-ID.** A globally unique identifier assigned to every message by the sending server. **In-Reply-To and References.** Headers that point at the parent message and the whole ancestry. Together they are how clients reconstruct a conversation, per [how email threading actually works](/blog/email-threading-for-agents). **Thread.** Not a thing the protocol stores. A chain assembled from those headers, plus each client's own guesswork. **Quoted history.** The copy of previous messages appended to a reply. Expensive in tokens and useless to a model, which is why it is stripped. **MIME.** The format that carries multipart bodies, attachments, and encodings. What raw email actually looks like, and what you do not want a model reading. **Envelope sender.** The address used in the SMTP transaction, often a provider bounce domain, distinct from the From header a person sees. **List-Unsubscribe.** A header providing one-click opt-out, now required on bulk mail by the large providers, per [the requirements Gmail and Yahoo now enforce](/blog/bulk-sender-requirements). ## Agent and tooling terms **MCP.** Model Context Protocol, a standard letting an agent client discover and call tools from a server, per [what an MCP email server is](/blog/mcp-email-server). **Tool.** A function the model can call, with typed parameters and a description it reads as instruction, per [designing the tools your email agent calls](/blog/designing-tools-for-email-agents). **Grounding.** Requiring every factual claim to trace to something the agent actually read, per [making an agent answer only what it can prove](/blog/grounding-and-citation-email-agents). **Prompt injection.** Text in an untrusted input that the model treats as instructions. An inbox is the most exposed surface for it, per [prompt injection by email](/blog/prompt-injection-email-agents). **Idempotency.** Handling the same event twice without doing the work twice. Non-negotiable because webhook deliveries retry. **Escalation.** Handing a thread to a person. Should be a real tool the model can call rather than a behaviour described in prose. ## Product terms you will meet here **Identity.** One agent: a handle, a mailbox, and a key. The unit that scope, attribution, and revocation attach to. **Reply versus cold send.** A reply goes to someone who wrote first in that thread. Anything else is a cold send, which carries an unsubscribe link and is watched by the duplicate-content and velocity tripwires. **Allowance.** The daily and monthly send budget, drawn down by replies and cold sends alike. **Tripwire.** An automatic check on cold sends: duplicate content across recipients, or a sudden burst. **Grant.** Explicit permission for one agent to reach a note, a secret, or another identity's mail, per [deciding what each agent is allowed to see](/blog/agent-permissions-and-grants). **Refusal reason.** A stable string returned when a send is rejected, such as `daily_send_quota_exhausted` or `recipient_suppressed`, written for an agent to act on rather than to log. ![The same message described in three vocabularies](/blog/email-agent-glossary/layers.webp "One message, three vocabularies: transport, protocol, and agent.") ## The three that get confused most **Delivered is not placed.** A provider accepting a message says nothing about which folder it landed in. **Passing is not aligning.** SPF and DKIM can both pass on a domain that is not the one in From, which is a DMARC failure. **Transactional is not conversational.** A receipt expects no answer; an agent's message does, and they want different infrastructure, per [transactional, marketing, conversational](/blog/transactional-vs-conversational-email). ## Questions ### What is the difference between SPF, DKIM, and DMARC? SPF lists authorised servers, DKIM signs the message cryptographically, and DMARC decides what happens when neither aligns with the visible From domain. DMARC is the one that determines the outcome. ### What does alignment mean? That the domain which passed SPF or DKIM matches the From domain a recipient sees. Without it, both checks can pass while DMARC fails. ### What is a cold send? Any message to someone who did not write first in that thread. It carries an unsubscribe link and is watched for duplicate content and sudden volume. ### What is MCP? Model Context Protocol: a standard for an agent client to discover and call tools from a server, so email becomes a URL and a key rather than an integration project. ### What is grounding? Requiring that every factual claim in a reply traces to something the agent read in that run, and escalating when it cannot. ### Why does idempotency keep coming up? Because webhook deliveries retry by design. Without a handled-message check, an agent will eventually reply twice to the same message. ## Related - [How to give an AI agent its own email address](/blog/email-for-ai-agents) - [SPF, DKIM, and DMARC for an agent subdomain](/blog/dmarc-record-generator-agents) - [What an MCP email server is, and how to connect one](/blog/mcp-email-server) - [Documentation](/docs)