We Gave Our Website an Agent — and an A2A Endpoint
Atakumi · 2026-08-23
Short answer
Atakumi's website ships with two agent surfaces built on one knowledge core: /talk, a full-page ChatGPT-style interface where people can ask anything about how we work, and /.well-known/agent.json, an A2A-protocol agent card that lets other AI systems query us directly. Both answer from curated, published material — and both point to a human when it matters.
Most agency websites have a contact form. Ours has a conversation.
As of this launch, atakumi.com answers questions about itself in two ways — one for humans, one for machines. This field note explains what we built and why.
/talk — the human surface
Head to /talk and you get a full-page conversational interface, in the family of ChatGPT or Claude — not a chatbot bubble in the corner. Ask what a pathway sprint involves, what an engagement costs, what you own when we leave, whether AI even makes sense for your kind of business.
The agent answers from a curated knowledge core: our published positioning, the pathway model, our essays, our FAQ. It doesn't improvise claims, it doesn't invent client names, and when the conversation turns to "should we work together", it hands off to a human booking — because that's a conversation we want to have ourselves.
agent.json — the machine surface
More interesting: /.well-known/agent.json. That's an A2A (Agent2Agent) protocol agent card — a machine-readable declaration that other AI agents can discover and query.
When a procurement team's AI assistant is asked "find me AI agencies and compare how they work", it can interrogate our agent directly: what does Atakumi do, how do engagements work, what's the governance approach. Ours answers, from the same knowledge core, with the same guardrails.
As far as we can tell, we're the first AI agency in our market to ship one. That won't be true for long — and that's fine. Being early to agent-readable business presence is a bet on where discovery is going: from search engines, to answer engines, to agents talking to agents.
Why we built it
Three reasons, honestly stated:
It practises what we sell. We tell clients that AI capability should be owned, governed, and embedded in real workflows. Our own site now demonstrates all three — including an AI system with a written scope, curated knowledge, and known failure modes.
It's honest marketing. An agent that answers "what does it cost?" with "scoped per scenario, here's why, book a call" is more useful than a pricing page performing false precision.
It's a field note worth reading. You're reading it. That's the content engine working as designed: build interesting things, publish what we learn, let the right clients find us.
The guardrails, briefly
The agent only answers from published Atakumi material. It declines off-topic questions. It never fabricates client names or results. Conversations aren't retained beyond the session. And the whole thing is documented — because we wrote an essay about why governance is the enabler, and it applies to us too.
Try it: talk to the agent. Or, if you prefer your intelligence biological: book a call.
Questions about how this applies to your organisation?
