ทวีตที่ปักหมุด
The harness is not incidental. It IS the control substrate.
If you want an LLM to actually control systems, you need what classical control theory assumes implicitly but agent frameworks skip entirely:
State — not chat history. A typed canonical representation: current observation, belief state, memory, latent variables / inferred regime, goal stack, pending actions, recent disturbances, confidence / uncertainty. Without this, there's no feedback. Just open-loop prompting.
Action space — not "free-form do anything." Typed, constrained actions with schemas and affordances: actuators, APIs, skills, prompts to sub-agents, code patches, browser actions, controller parameter updates. Constraints are what make control possible.
Transition feedback — every action must return structured consequences: success / failure, resulting state delta, uncertainty, latency, cost, side effects. No structured feedback = no closed loop.
Cost / reward — an explicit objective function or utility model: task completion, safety margin, cost, latency, resource use, human preference, reversibility. Without this, the agent optimizes for nothing.
Verifier / observer — an independent mechanism to evaluate the resulting trajectory. The controller cannot be its own judge. You need a separate evaluation loop.
Safety filter — like a barrier function layer, but for agent actions: forbidden tool calls, resource limits, policy bounds, rollback conditions, invariant checks, human-approval gates when necessary. This is the runtime shield.
This is why "prompt + tools" feels brittle. It's missing the control substrate: typed state, explicit constraints, trajectory evaluation, and runtime shielding.
The recursive self-improving loop needs this foundation. The harness IS the controller.
English
