What is loop engineering — and should you actually build one?
Loops beat one-shot prompts: inbound → skill → tools → Ask → send → wait. Email is a natural loop surface.
Discuss this post in AI
Send a pre-filled prompt to ChatGPT, Claude, Gemini, or Perplexity — get a summary, ask follow-ups, or compare ideas from this guide.
Prompt engineering optimizes a single completion. Loop engineering designs the repeated cycle that makes an agent useful across days: something arrives, the role acts under policy, a human gates risk, the world changes, and the system waits for the next signal. If you only craft prompts, you build demos. If you craft loops, you build operations.
Loops versus prompts
A prompt is a function call with text. A loop is an operating rhythm with state, timers, webhooks, and permissions. Neuro OS encodes that rhythm as roles, skills, sandboxes, connectors, and Ask rules. The model is replaceable; the loop is the company asset.
The email-shaped loop
Email is an ideal surface because the protocol already has inbound events, durable threads, and outbound actions:
- Inbound — message or forward lands in Agent Inbox
- Skill — classify and plan inside the role
- Tools — extract attachments, read 1C/Bitrix/amoCRM via connectors
- Ask — human approves send or external write
- Send — reply or forward
- Wait — webhook for the next message
That is the same pattern as the definitive guide, stated as an engineering discipline.
Four conditions before you build a loop
Build a loop when all four hold:
- Recurring work — volume repeats weekly, not once
- Clear ownership — a human owns the role and exceptions
- Gated side effects — sends and writes can wait for Ask
- Observable state — threads, logs, and memory can be audited
Skip a full loop when the job is a one-shot analysis with no outbound channel. A prompt may be enough. Do not pretend a chat tab is a loop because you re-pasted context by hand.
Should you build one?
Yes, if agents must stay in conversation with the business—support, collections, procurement, coding agents receiving OTPs. No, if you only need a sporadic draft and a person will always paste the result into Outlook themselves. Partial loops are fine: draft-only with Ask still beats silent auto-send.
Failure modes
- Auto-send without Ask burns trust and deliverability
- No idempotency doubles actions on webhook retry
- Prompt-only memory forgets the thread—use email as memory
- Missing identity stalls autonomy—email as identity
Product posture
Agent Inbox is API-first for agents; humans remain on Ask. Neuro OS keeps the loop in git so you can review skill changes like code. Self-host when policy requires. Loop engineering is not a slogan—it is the decision to treat agents as long-lived workers with channels, gates, and wait states.
If your “agent” cannot wait for mail, it is not ready for email work. If it can, build the loop deliberately—build email agents.
Minimal viable loop
You do not need a dozen microservices to start. You need: an inbox, a webhook into Neuro OS, one skill, Ask on send, and logs you can read later. Add classifiers, allow lists, and multi-tenant splits as volume demands. Many failed agent projects skip the wait state and try to finish the job in one prompt.
Organizational design
Loops need owners. Assign a role owner, an on-call for stuck Asks, and a weekly review of exception themes. Without that, Ask becomes an ignored queue and people bypass the agent. Loop engineering includes the human staffing model, not only the graph of tools.
Agent Inbox gives each role a mailbox people can reply to, with forwarding into Neuro OS when a human must see the thread. Outbound mail defaults to Ask. Run the role on Neuro OS. To scope the first inbox, get started.