System Architecture Diagram

Draft · updated 2026-10-30 · embodied-ai-architect/docs/plan/roadmap-system-architecture.md

The System Architecture Diagram visualizes how the “Agent” moves from an abstract reasoning entity to a practical engineering tool that manipulates other software.

Here is the architecture diagram of our Agentic HW/SW Co-Design Platform.

System Architecture Diagram: The Agentic Co-Design Loop

This diagram illustrates the iterative loop where the Orchestrator Agent acts as the central nervous system. It takes high-level human constraints, generates concrete HW/SW designs, tests them in a “Virtual Lab” (compounded simulators), and uses the resulting energy and performance data to reason about the next design iteration.

graph TD
    %% --- Actors and Initial Inputs ---
    User([Human Architect]) -->|Define Goals & Constraints| Requirements[Constraints & Objectives Interface]
    Requirements --> Orchestrator

    %% --- The Core AI Agent ---
    subgraph AgentCore ["Agentic AI Core (The Architect)"]
        Orchestrator[Orchestrator Agent LLM Reasoning & Planning]
        Memory[(Design History & Knowledge Base)]
        Orchestrator <-->|Read/Write Past Experiments| Memory
    end

    %% --- The Design it is Manipulating ---
    subgraph DesignState ["Candidate Design State"]
        Orchestrator -- "Selects/Tunes Components" --> SW_Design[Software Pipeline Config]
        Orchestrator -- "Parametrizes Hardware" --> HW_Design[Virtual HW Description]
    end

    %% --- The Simulation Infrastructure ---
    subgraph SimLab ["Virtual Testing Lab (The Tool Layer)"]
        Mission[Mission Scenario Definition] --> CoSimEnv
        HW_Design --> CoSimEnv
        SW_Design --> CoSimEnv

        subgraph CoSimEnv ["CoSim Environment"]
            direction LR
            FlightSim[Physics Simulator] -- "Sensor Data" --> ComputeSim
            ComputeSim[Compute Simulator] -- "Control Commands" --> FlightSim
        end

        ComputeSim -- "Cycle-Accurate Activity Traces" --> PowerModel[Power Estimator]
    end

    %% --- The Output and Feedback ---
    subgraph Feedback ["Feedback Analytics Layer"]
        FlightSim -->|"Logs (Crashes, Time)"| Metrics_Phys[Physics Metrics]
        ComputeSim -->|"Logs (Latency, Bottlenecks)"| Metrics_Perf[Performance Metrics]
        PowerModel -->|"Logs (Joules, Watts)"| Metrics_Energy[Energy Efficiency Profile]

        Metrics_Phys --> Aggregator[Feedback Aggregator]
        Metrics_Perf --> Aggregator
        Metrics_Energy --> Aggregator
    end

    Aggregator -- "Agent Reads and Reasons" --> Orchestrator

    %% --- Final Output ---
    Orchestrator -- "If Objectives Met" --> FinalBlueprint[Finalized HW/SW Blueprint Package]

Diagram Breakdown for Markdown Documentation

1. The Agentic AI Core (The Architect) At the center is the Orchestrator Agent, an LLM-based system capable of complex reasoning. Unlike simple optimization scripts, it doesn’t just tweak numbers; it understands architectural concepts. It relies on a Design History & Knowledge Base (Memory) to remember past failures (“Reducing L2 cache too much causes SLAM tracking loss”) and applies that knowledge to future iterations.

2. The Candidate Design State In every iteration, the Agent outputs a specific, testable configuration. This is split into the Software Pipeline Config (e.g., deciding to use an int8 quantized YOLO model running at 15fps) and the Virtual HW Description (e.g., defining a RISC-V core with a specific custom accelerator block attached).

3. The Virtual Testing Lab (The Tool Layer) This is the engine room. The Agent doesn’t run simulations itself; it uses “tools.”

• Co-Simulation Environment: This crucial block tightly couples physical world simulation (e.g., AirSim generating camera feeds) with digital hardware simulation (e.g., qemu/csim executing the software stack on virtual silicon). They must run in lockstep to ensure that a slow hardware design actually causes the drone to react slowly to obstacles in the physics simulator.
• Power Estimator: It takes the detailed activity logs from the compute simulator (recording exactly which transistors were switching) and translates that into energy consumption metrics (Joules/Watts).

4. The Feedback Analytics Layer Raw logs are turned into actionable insights. The output isn’t just “it took 100 Joules.” The output is a structured profile indicating, for example, that “The perception neural network consumed 60% of the total energy budget, and it was memory-bandwidth bound, not compute-bound.”

The Iteration Loop The Feedback Aggregator feeds this structured profile back to the Orchestrator Agent. The Agent reads this “report card,” reasons about the bottlenecks, and begins the loop again by modifying the design state for the next test.