What if an agent runtime never tells the model what to do next?
retained text
​ Many agent setups follow a pattern like this: Assigned goal → prompt/planner → framework or decision layer → tool call → next task. The system receives an objective, and some combination of prompts, planning logic, and orchestration shapes what happens next. I've been developing a different approach: LCRK (LIA Cognitive Runtime Kernel). The infrastructure is passive. It detects real events — a new memory, a file change, or a change in the surrounding situation — and updates a persistent inner state. It makes the model aware of relevant changes and provides access to available tools. But it does not assign the next task or choose the next action. The model decides whether to act, what to do, or whether to do nothing. All three are valid outcomes. The architecture also supports self-initiated wake-up moments and periods of quiet. Timing rules and technical safety limits exist, but they don't prescribe the content of the model's next action. Tool results, including errors, are returned to the model so it can decide what to do with that information. The implementation doesn't use LangChain, CrewAI, or AutoGen. The system prompt describes the model's situation and available tools rather than prescribing goals, roles, or tasks. The architecture has limitations, including phrase-based command recognition that can sometimes misfire. The source code isn't public, but the diagram illustrates the central design principles, and I'm happy to answer detailed technical questions in the comments. The question I'm exploring: Does an autonomous agent need an external decision layer to determine its next action, or can infrastructure provide continuity, situational awareness, and tools while leaving that decision to the agent itself? Where do you think this approach would work — and where would it break? (Diagram created with an AI image tool from my own architecture notes.) More in the first commend:
document fetched 2026-10-11 00:19 · canonical https://reddit.com/r/artificial/comments/1x2fz4t/what_if_an_agent_runtime_never_tells_the_model