Architecture
15 posts on architecture for AI agents that send and receive email.
6 August 2026 · 4 min read
The forwarded message problem
A forward carries a stranger's words, a broken reply-to, and instructions the agent should not follow. How to parse one, who to answer, and why forwards break authentication.
5 August 2026 · 4 min read
You already send with a transactional provider. Now it has to receive
Resend, Postmark, SendGrid, and Mailgun send well. The moment an agent has to read replies you need threading, contact memory, and an inbox, which is a different product.
5 August 2026 · 4 min read
Building the agent's mail layer yourself on SES
SES is cheap and capable, and an agent needs a dozen things it does not provide. An honest list of what you would build, how long it takes, and when doing it is the right call.
5 August 2026 · 4 min read
Where an agent fits if you already run a helpdesk
You already have a ticketing system and it is not going anywhere. Four ways to add an agent around it, which one to pick, and how to avoid two systems answering the same customer.
4 August 2026 · 4 min read
Designing the tools your email agent calls
Most agent failures trace to tool design rather than the model: too many tools, vague names, errors that invite retries, and results that hand the model raw text. What good looks like.
3 August 2026 · 4 min read
When an agent should send one email instead of nine
Agents can send a message per event, and usually should not. How to decide between immediate, batched, and digest delivery, and how to implement the batching without losing urgency.
2 August 2026 · 5 min read
Where each piece of an agent's state belongs
Thread, contact, note, vault, your database, and the agent's own storage each hold a different kind of state. Putting something in the wrong one is why agents drift, repeat themselves, and contradict your systems.
2 August 2026 · 5 min read
Giving every customer their own agent inbox
If your product runs an agent for each customer, each one needs its own address. How to provision at signup, keep tenants isolated, handle domains, and keep one tenant's mistake from reaching the others.
31 July 2026 · 5 min read
What happens when forty messages arrive at once
Agent email is bursty. How to handle a queue of inbound mail without double-replying, racing on one thread, or spending your daily allowance in ninety seconds.
28 July 2026 · 5 min read
Moving an agent off the Gmail API onto its own inbox
A migration plan for agents built on Gmail or Graph: what breaks at scale, how to move one flow at a time without a cutover, and what to do about history, tokens, and the old address.
28 July 2026 · 5 min read
How email threading actually works, and why agents break it
Threading is a convention held together by three headers and a lot of client guesswork. What they do, how agents break conversations, and why a broken thread is worse than a wrong answer.
27 July 2026 · 5 min read
An agent that signs itself up for its own inbox
Autonomous agents need addresses before a human is available to create one. How self-signup works, what a claim URL is for, and the guardrails that make agent-provisioned inboxes safe to allow.
26 July 2026 · 5 min read
Webhooks or polling: how your agent should learn that mail arrived
A practical comparison of webhook-driven and polling agents, with signature verification, retry and idempotency handling, and the hybrid most production deployments end up running.
26 July 2026 · 5 min read
Long-running email conversations: waiting, timeouts, and resuming
Email conversations take days; agent runs take seconds. The patterns for holding a thread open, choosing timeouts, resuming with context, and knowing when to stop following up.
26 July 2026 · 5 min read
Running several email agents in one workspace
How to structure a workspace once you have more than one agent: address naming, what to share and what to isolate, handoffs between agents, and the failure modes that only appear at three or four.