Start with something human
01 / A simple requestA clear sentence. Not yet a clear instruction.
An illustrative scenario. No AI inference, calendar connection or real action.
We speak with shared history, relationships and expectations. Much of what we mean never makes it into the sentence.
An agent needs enough of that meaning to know what to do, what to respect, and when to ask. More tools alone cannot fill that gap.
Konnyaku
An architectural directionWe’re exploring how human meaning can become explicit enough for an agent to act appropriately—or recognize that it should wait.
What I mean.
Enough clarity to act. Or a reason to ask.
The intended output: action-ready understanding, with unresolved questions made visible. Execution remains with the agent.
How should we ask the model?
What does this data mean?
What does this person mean in this context?
These questions overlap. Prompts, data semantics, retrieval and memory can all contribute. Konnyaku’s focus is interpreting intent before execution—not replacing those foundations or translating between languages.
Konnyaku does not replace agents. The proposed layer helps clarify what they are acting for.
Intent · Language · Context · Preferences
Interpretation · Context resolution · Constraints · Meaning representation
If something is missing: ask, defer or stop.Reasoning · Tools · Workflows · Execution
Applications · APIs · Data · Services
A proposed architecture, not a list of available integrations. Sources of context, permission checks and system connections would need to be established and validated in each implementation.
As agents move from answering to acting
A wrong answer can mislead. A misunderstood instruction can change a record, send a message or move someone’s money.
Understanding more should not mean taking more control. Sometimes the right outcome is: do not act yet.
“Move my meeting with Alex to tomorrow afternoon.”
The name draws on the familiar idea of “translation konnyaku”: understanding one another without first learning a different language. It brings that imagination to a different question—what would it take for technology to understand human intent?