Agent Operational Cognition

Definition

Agent operational cognition is an agent’s situated understanding of how an organization actually performs work: its rules, goals, workflows, exceptions, approvals, task history, and reasons for acting.

Current Synthesis

The source separates operational cognition from interface execution with a “brain” and “hand” analogy. A system may click, type, browse, and manipulate software competently while still misunderstanding what the business task means or which local rule governs it. Reliable enterprise agency therefore requires computer use to be joined with accumulated organizational context, feedback, memory, and judgment.

Key Claims

• Interface control and business understanding are distinct capabilities.
• Software execution is necessary because digital work ends in applications, websites, files, or APIs.
• Organizational rules and exceptions determine what action is appropriate, not only whether the action can be executed.
• Repeated work and feedback can build situated competence that a generic model does not possess initially.
• Stronger base models may simplify execution scaffolds without eliminating local cognition or verification needs.

Evidence

Brain-hand distinction

• 【年终复盘②】RSI可能是NeoLab的机会、PA真的有用、创业要造灭火器还是砖头？｜对话谷雨 says software and web operation are the “hand,” while understanding company process and task intent is the “brain.”

Learning local work

• 【年终复盘②】RSI可能是NeoLab的机会、PA真的有用、创业要造灭火器还是砖头？｜对话谷雨 uses repeated tasks, feedback, conventions, workflows, and approval rules to define the missing cognition.

Scaffolding change

• 【年终复盘②】RSI可能是NeoLab的机会、PA真的有用、创业要造灭火器还是砖头？｜对话谷雨 reports that stronger models reduced the need to sample many trajectories, select a correct one, and package it as a reusable skill in every case.

Counterevidence & Qualifications

• The claim that computer-use execution is nearly solved conflicts with unresolved reliability, speed, permission, privacy, rollback, and verification evidence elsewhere in the wiki.
• Organizational cognition can become stale or encode harmful local practice; learning requires provenance, correction, access control, and deletion paths.
• Better task success does not reveal whether the agent formed transferable understanding or merely exploited familiar patterns.
• The source provides operator judgment rather than comparative benchmark evidence for the brain-hand split.

What Changed

• Established operational cognition as the organizational-knowledge layer distinct from interface manipulation.

Related Concepts

• Computer Use Agent - execution layer whose capability does not guarantee correct task understanding.
• Enterprise Operational Memory - retained organizational context supporting situated action.
• World Models - broader representation of environments in which agents act.
• Continual Learning - mechanism for acquiring and revising local competence.
• AI Skills - reusable procedures that can encode part of a workflow.
• Agent Reliability Verification - checks whether understood intent became correct execution.