Skip to main content
Adsum IoT Coder is an AI coding agent built on the open-source Cline foundation, with IoT-specific knowledge modules and tool-use skills layered on top.

​
Modules loaded on demand

Domain knowledge and tool-use skills are structured as a framework of discrete, composable modules, each scoped to a specific chip family, protocol stack, or debug capability. At session start, the agent assesses what the project is and what the task requires, then loads only the relevant modules. 
User task ──► Agent assesses project ──► Loads scoped modules ──► Executes
              (chip family,               (only what's needed     (Build → Flash →
               protocol stack,             from iot-knowledge/)    Capture → Analyze
               debug category)                                      → Fix, looped)
The module tree on disk: 
iot-knowledge/
├── rules/                  # Platform-agnostic agent constraints
├── platforms/nrf/          # Nordic nRF (NCS / Zephyr)
│   ├── boards/             # Per-SoC specifics
│   ├── sdks/ncs/           # Project structure, Kconfig, BLE stack
│   ├── workflows/          # Entry-point sequences (start here)
│   └── actions/            # Subroutines loaded by workflows
└── platforms/esp/          # Espressif ESP (ESP-IDF)
Analyzing a UART log drop loads the log-analyzer workflow plus the capture-logs action and the NCS SDK module. Debugging a failed BLE connection on a two-board setup also pulls in the BLE module, the device-identity rule, and the relevant board file, and nothing else. The model gets exactly what the task requires, no more.

​
Why it matters: context quality

The payoff is not just avoiding context overflow. Even when a full static bundle would technically fit, loading only the relevant modules keeps domain knowledge in the model’s effective working set rather than letting it get buried as the session grows. This is the “lost in the middle” failure mode the benchmark caught the general-agent baseline hitting: same model, full window, lost the original symptom by the fourth debug cycle.

​
Human-curated, not AI-generated

A common trend in AI tooling is letting agents author and refine their own tool-use skills. Our own research and experimentation led us in the opposite direction for high-stakes IoT debugging. The knowledge that drives Adsum is hand-authored or hand-reviewed by senior IoT engineers. AI-generated skills can read fluently and still encode subtle misunderstandings of a protocol stack, failing in ways the agent cannot self-detect. Expert curation is the bottleneck that keeps the quality bar honest. This direction is where frontier research points too: equipping a general model with curated, domain expertise that loads only when needed, rather than scaling the model alone. See the academic work on expert-skill-augmented models that shaped our benchmark (arXiv:2603.19583) and industry practice on context engineering and Agent Skills.