Most guides to building AI agents are written from theory. These were written from production: what we learned running AI employees across millions of live WhatsApp conversations, where a single wrong reply is a real customer, not a benchmark. Twelve principles, the standard we hold every build to, and how it all maps to the wider field.
Our principles come in two tiers. The first governs orchestration: how several agents share one conversation without talking over each other. The second governs conduct toward the customer: truth, safety, and their rights. The first tier is what most teams get wrong at the mechanics level; the second is what they forget entirely.
A language model is trained to be helpful, which means every agent, left alone, tries to answer everything. When you run more than one agent in a conversation, that instinct produces chaos: two agents answering the same question, often with different numbers. These five rules exist to make not answering feel correct to the model.
At any point in a conversation, exactly one agent speaks. If the router is routing, it does not also answer. If a specialist is answering, no other agent is posting.
In the field: Google's conversation-design guidance calls this turn-taking; OpenAI's Agents SDK enforces it structurally as a single active agent per exchange.When an agent hands off, it says nothing. No "I'll connect you", no "one of our team will help". The receiving agent is responsible for the next customer-visible message, and picks up as if it had been there all along.
In the field: a cleaner version of the handoff primitive in OpenAI's Agents SDK, and Grice's maxim of Manner: say no more than the moment needs.An agent greets and introduces itself only if it is the first agent in the thread. If another agent has already spoken, the next one continues directly. No "Hi, I'm the sales assistant" three messages in.
In the field: Microsoft's Guidelines for Human-AI Interaction, guideline 12: remember recent interactions.Every agent carries a short list of what it answers and a short list of what it routes. Scope is declared in writing, never left for the model to infer in the moment.
In the field: Anthropic calls this routing with separation of concerns; Microsoft's guideline 10: scope services when in doubt.When in doubt, the router routes and a specialist escalates to a human. The model's default is to help by answering. We design against that instinct explicitly, because a confident wrong answer costs more than a handoff.
In the field: Grice's maxim of Quality, never assert what you cannot support, expressed as an orchestration rule.Here is the failure these five prevent. We call it the double-post: the router answers, then hands off to a specialist who answers again, and the customer gets two replies with two different prices.
The router answered instead of routing, Sales re-introduced itself, and the two prices contradict. The customer feels the machinery.
One agent takes the whole turn. The customer sees a person who knows the business, not a relay of bots.
The first tier keeps agents from tripping over each other. The second keeps them honest, safe, and answerable to the person on the other end. These are the principles most "AI agent" pitches skip, because they are invisible until the day they are not.
Every material fact, a price, a date, an availability, a policy, comes from the business's own knowledge base or a human. If it is not there, the agent says so and routes. An agent that invents to stay helpful has failed, even when the customer never notices.
In the field: Grice's maxim of Quality, and the trust and explainability chapter of Google's People and AI Guidebook.Our agents never adopt a human name or pretend to be a person. Ask one directly whether it is AI and it tells you, in one line, and carries on. Honesty about what it is costs nothing and keeps the whole conversation trustworthy.
In the field: the Nielsen Norman Group's first rule for chatbots, disclose the bot; and the direction the EU AI Act sets for AI disclosure from August 2026.Every deployment has a real, named escape hatch to a person, and the agent takes it the moment a customer asks or the situation needs judgment. An AI with no exit to a human is a trap, not a service.
In the field: the Nielsen Norman Group names "no path to a human" as a core chatbot failure; Microsoft's guideline 17: provide global controls.Before any booking, payment, reschedule, or other change it cannot take back, the agent repeats the exact detail and waits for a yes. A wrong read costs a correction. A wrong write costs a customer.
In the field: Microsoft's guideline 16: convey the consequences of user actions.The agent never requests or stores sensitive data in chat, card numbers, full identity numbers, passwords, and never re-asks for something the business already holds on record. Every field not asked for is a field that cannot leak.
In the field: data minimisation, as set out in Google PAIR's data guidance and ISO/IEC 42001.Every live deployment is read against defined numbers every week: how often agents double-post, re-introduce, drop a handoff, escalate a false alarm, or contradict themselves. A principle you do not measure is a preference.
In the field: Anthropic's guidance to optimise with evaluation and add complexity only when it demonstrably helps; Microsoft's guideline 2.When a system is down, a lookup times out, or a question falls outside scope, the agent degrades to a clean human handoff. The customer sees a helpful pause, never an error message, and never a guess dressed as an answer.
In the field: the errors and graceful-failure chapter of Google's People and AI Guidebook.
Principles are direction. This is the specific, written standard every AI agent we build carries, sitting inside each agent's instructions where the client can read it, not hidden behind a platform setting. If it is not on this page, we did not commit to it.
These apply to every agent on a deployment: the sales agent, the booking agent, the support agent, and any other specialist we build.
The rules above are what agents are committed to. This is what AI, as a technology, cannot do, and we would rather say it plainly than sell around it.
Despite all of the above, agents drift. They meet an edge case we missed, quote a stale policy, or misread an intent and route the wrong way. When that happens, here is what we do.
The things every active deployment receives, without exception, for the life of the engagement.
At least twenty real threads read end to end, every week, by a senior operator, with issues flagged and logged.
We run our QA script before shipping any prompt or knowledge-base update to a live agent. Nothing changes silently.
A short end-of-month summary: what we fixed, what we are watching, what is trending in your conversations. Two-minute read.
Every change dated, described, and attributed. Request the full log any time.
When something breaks, or when the standard itself needs to change, you reach us directly. You are never routed to a ticket queue. A founder is always your escalation path.
We did not invent these principles from a textbook. We arrived at them by running agents in production and watching what broke. But we are not the first people to think about how machines should hold a conversation, and honest work shows its sources. There is no single agreed standard for AI agent design the way there is for web accessibility. There is a body of serious thinking, and our principles converge with it, and in the orchestration details go past most of it.
We design, deploy, and run AI employees that follow every principle on this page: scoped, grounded, measured, and answerable to a human. Tell us what your customers keep asking.
Build your team