The future of deliverability: agents, not dashboards

The deliverability industry spent ten years building prettier dashboards. Nicer charts. More filters. Exportable PDFs. All of it aimed at the same user: a human marketer who logs in once a week, scrolls, and maybe opens a ticket. That user is becoming a minority.

In 2026 the fastest-growing consumer of deliverability data is an AI agent. Not a scheduled cron job — a stateful assistant that discovers tools, runs them, reasons over the output, and acts. The implications for anyone building in this space are larger than most people realise. The dashboard is not going away, but it is moving to the edge of the product instead of the centre.

The old pattern: human → dashboard → fix

For the last decade the loop was:

1. Deliverability problem appears (opens drop).
2. Human notices in the weekly report.
3. Human logs into GlockApps / Mail-Tester / Inbox Check.
4. Human reads a chart.
5. Human copies a DNS recommendation into a ticket.
6. A DevOps engineer eventually changes the record.

The feedback cycle is measured in days. Half the tickets never close. Every vendor competes to make step 4 easier to read, because steps 1, 2, 3 and 5 are outside their control.

The new pattern: agent → tool via MCP → action

Swap a human for an agent and the loop collapses:

1. Agent watches an engagement stream.
2. Drop detected → agent calls start_test via MCP.
3. Agent reads verdict, cross-references DNS, forms a hypothesis.
4. Agent proposes a Cloudflare / Route53 change as a pull request.
5. A human approves the PR.

Cycle time: minutes. Steps that used to sit in four different tools, three different dashboards and two different teams now live in one conversation. The human is still in the loop — at exactly the point that matters (approving the change) and not the points that don't (scrolling a chart).

Why A2A matters

Agents rarely work alone. A deliverability agent finds a bad SPF record; fine, but who fixes it? Probably a DNS-management agent, which lives at a different vendor and speaks a different protocol. Without a standard for agent-to-agent communication, every pairing needs a bespoke bridge. A2A is that standard — the same role HTTP played for web services in 1999.

The value is not in any single agent. It is in the network effect of a hundred agents that can discover each other via .well-known/agent.json, negotiate capabilities, and hand work off. The deliverability vendor that ships a first-class A2A surface in 2026 will be a default integration in 2027. The one that doesn't will be a tab people forget to open.

The vendor shift: from UI to SDK + MCP + A2A

Concretely, the product investment looks different:

What matters more

• Machine-readable API docs (OpenAPI, JSON Schema).
• MCP server that exposes the same surface as the API.
• A2A agent card at the well-known path — versioned, signed where possible.
• Structured output formats for every verdict.
• Webhooks for event-driven agents (faster than polling, cheaper than idle connections).

What matters less

• Custom dashboard theming, dark mode, export buttons.
• In-app notification centres.
• Hand-curated PDF reports (agents regenerate them on demand for each stakeholder).

This is not a speculation. You can watch it happen — the vendors advertising "MCP support" and "agent-ready" APIs in 2026 are the same ones growing developer signups 3–5× year on year.

What breaks: pricing by user seat

The easiest thing to miss: a seat-based pricing model assumes a single human per login. An agent can do the work of ten marketers, but it still counts as one seat. The obvious patch — "agent seats" — is unstable because one agent might call your API a thousand times more than a human ever would. Usage-based pricing is the only model that survives contact with autonomous workloads.

Expect most vendors in this space to migrate to metered pricing within 18 months. The ones that don't will either bleed money or throttle agents so hard they become unusable.

Three concrete predictions for 2027

1. MCP becomes table stakes

By mid-2027 any deliverability tool that does not ship an MCP server will look as outdated as one without a REST API looked in 2015. Buy cycles will include "do you support MCP?" as a hard requirement from DevOps.

2. Agents replace most junior DMARC analysts

The work of parsing RUA/RUF reports, correlating them with sending services, and proposing record changes is almost entirely automatable. A level-3 agent does this job in minutes. Senior analysts shift to strategic work — brand alignment decisions, sender identity architecture, vendor selection — where human judgement still wins.

3. Deliverability checks embed in dev tools

The test moves left. Instead of a placement test after deploy, your transactional-email template change triggers an agent run in CI, and the PR is blocked until the placement score is above your threshold. We already see design partners doing this today. By 2027 it will be the default setup in well-run companies.

Counterpoints

None of this means the human disappears. A few things do not automate well, and probably will not for years:

• Brand-level decisions. Should a new product launch use the corporate domain or a subdomain? That is a strategic call, not a DNS one.
• Relationship work. Talking to a Gmail postmaster contact, filing a mitigation request, building a direct peering relationship. Humans do this better.
• Novel failure modes. When a brand-new ML filter rolls out at a major provider, the first week is diagnostic chaos. Agents lag the curve by weeks because their training data does not yet include the new signal.

Frequently asked questions