A fast-agent layer that helps humans compose more productive messages to high-context LLM agents from low-bandwidth interfaces.
The decoder side of the agent-human I/O pair. Counterpart to MOP — Model Output Protocol.
When you drive coding agents from a phone (Telegram, voice dictation, no keyboard, no terminal), the constraints flip:
- Latency is high. Typing on glass is slow. Voice is faster but lossy.
- Observation surface is small. You can't see the codebase, the logs, or the agent's recent work.
- Control surface is small. No keyboard shortcuts, no tab-complete, no file picker.
Result: messages are terse, ambiguous, miss context the agent would benefit from. The agent burns expensive turns guessing what you meant or asking for clarification it could have figured out.
HOP sits between the human and the agent in the input direction. A fast model (Haiku) processes the human's message before the slow agent sees it:
- Dictation cleanup — normalize voice transcription artifacts.
- Intent expansion — resolve ambiguous references ("that thing we talked about") against recent context.
- Context preloading — based on detected intent, attach relevant skills, memory, or file refs to the agent's context window.
- Pre-flight ambiguity check — if the message is incomplete or contradictory, surface it BEFORE the slow agent burns turns guessing.
- Composition assist — fast agent suggests a clearer phrasing with one-tap accept.
The slow agent receives a richer, less ambiguous prompt. Fewer wasted turns. Faster time-to-action.
Same modes pattern as MOP, configurable per-chat:
- Passthrough — input flows through unchanged.
- Audit — log enhancement opportunities but pass original through.
- Suggest — surface a rewrite for one-tap accept; original sends if ignored.
- Auto — apply enhancements transparently, log original + enhanced.
Start in Audit. Watch what the fast agent would have done. Promote when trust builds.
Design phase. Architecture and rule format defined; implementation follows MOP's foundation.
MOP — Model Output Protocol. The encoder side: filters model output for human bandwidth limits.
Together, MOP and HOP form the I/O contract for the agent-human interface. Like Swift's Codable, but for human bandwidth limits.
MIT