Everyone’s building AI agents. Most of what gets called an “agent” is a chatbot with a longer system prompt.

An actual agent does three things a chatbot cannot: it plans, it acts, and it remembers. Not in a metaphorical sense — in a mechanical one.

Planning means structured decomposition

A chatbot responds to a prompt. An agent receives a goal and breaks it into steps it can execute in sequence, adjusting the plan when earlier steps produce unexpected results.

The key word is adjusting. If step two fails, a proper agent revises the plan rather than returning an error. That’s not a feature you add with a clever prompt — it requires the architecture to support it.

Acting means calling real tools

An agent without tool use is still a chatbot. Acting means: searching the web, reading a file, writing to a database, triggering a webhook, sending an email. Real-world effects, not text about effects.

Tool use is where most implementations fall apart. The prompts that make tool use reliable — handling errors, retrying correctly, knowing when not to use a tool — are what the Shinpo agent kits are built around.

Remembering means context that persists

A chatbot knows only what’s in its current context window. An agent can store information, retrieve it later, and build up a picture of your business over time.

Without this, you rebuild context from scratch every session. With it, the agent compounds — every interaction makes it more useful.


The six yōkai in the Shinpo kit are each designed around one of these capabilities as a primary specialisation. Kitsune plans. Tengu acts. Kappa remembers. Oni guards. Tanuki adapts. Ryū orchestrates all of them.

That’s what makes them agents rather than chatbots with names.