Summary. My posts explore the capabilities of AI agents in executing complex tasks, from building games and hardware integrations to automating legal documentation. I examine how different agentic frameworks and tools handle ambiguity, verification, and the integration of local versus hosted models.

What the posts say

The agent successfully built the game, verified its own work through logic and headless-browser tests, and produced a fully functional product.

“DeepSeek Harness is a developer-preview tool where everything is a plugin — models, tools, sandboxes, storage, even the whole UI can be swapped or recomposed.” — DeepSeek Harness (Part 1)

The agent performed 96 Node logic checks to verify the game mechanics.

“96 Node logic checks covering spawn, movement, collisions, shield erosion, respawn and invulnerability, UFO scoring, wave transitions, the ground-line game-over, and restart.” — DeepSeek Harness (Part 1)

The agent automatically installed the required toolchain and libraries when it found them missing.

“Claude created the project folder, discovered arduino-cli was not installed, and worked out the toolchain from scratch: checking serial devices, installing arduino-cli via Homebrew, then the arduino:avr core and LiquidCrystal library.” — Claude Code (Part 15)

The agent created a local wiki to store knowledge about the hardware quirks.

“Along the way Claude also stood up a small .llmwiki in the project — its own notes on the board's real wiring and the sketch's design — so the next session starts from what was actually learned here rather than rediscovering the D12-is-really-D11 quirk from scratch.” — Claude Code (Part 15)

BMAD is the only tool that proactively asks for clarification on ambiguities.

“It is the only tool that actually asked. Every ambiguity I tracked across this series — bitwise/word-size scope, negative-number representation, sample-vs-population standard deviation — BMAD stopped and asked me directly, as real choices, rather than silently deciding (OpenSpec) or silently deciding with written rationale (Spec-Kit).” — Spec-Driven AI Tools (Part 3)

The three-layer review process caught several bugs before I saw them.

“Its three-layer review found real defects and correctly rejected several false findings.” — Spec-Driven AI Tools (Part 3)

The local model required more detail because a specification for a cloud model is too loose for a 30-billion-parameter model.

“A specification good enough for a frontier cloud model can be far too loose for a 30-billion-parameter local model.” — DeepSeek Harness (Part 2)

The hybrid model works if the tasks are small and clearly defined.

“The short answer: it works, provided the pieces are small and written down exactly.” — DeepSeek Harness (Part 3)