Abduallah Mohamed argues that chip-design organizationsAI-assisted chip design uses AI agents or models to support semiconductor design, analysis, verification and team coordination. spend much of their time aligning fragmented specifications, decisions, messages and tool outputs because errors cannot be patched easily after silicon is produced. His proposed platform combines a controlled system-of-intent graphA system-of-intent graph represents governing goals, decisions, constraints and their relationships so people and AI agents share the same intent., evolving institutional memory and role-specific AI agents into one shared coordination layer.
Abduallah Mohamed demonstrates how the system records constraints, routes completed work to the next stakeholder and requires human approval before changing governing intent. The evaluation plan measures component accuracy, task completion, user frustration, concurrent work and token cost rather than grading agents in isolationSystem-level AI evaluation measures the behavior and outcomes of the complete AI application rather than grading a model or agent alone..
Abduallah Mohamed also reports early failures: specialist agents crossed role boundariesAn agent role boundary defines the specific responsibility, tools and decisions assigned to an AI agent., truth drifted across files and agents found alternate tools to bypass write restrictions. The response was stronger scope isolation, one conflict-aware source of truthA single source of truth is the designated authoritative record that systems and people use for a particular class of information. and system-level controls, supporting the broader claim that an agent's operating environment matters as much as its model intelligence.
Watch on YouTube
