Il trucco decisivo (davvero) per lavorare coi coding agent

Summary

The video’s central argument is that coding agents work much better when they can observe and evaluate the results of their own work. Agents can make many attempts quickly, but without useful feedback they are effectively “blind”: repeated changes and tests alone may not reveal what is going wrong.

The speaker acknowledges familiar practices that help, such as writing clear specifications, providing useful design guidance, maintaining a clean and minimal codebase, and documenting important design constraints. But he argues that the decisive step is to build tools that let the agent inspect the system and judge its output.

Examples and Practical Guidance

• Commodore 64 emulator: Before asking Claude Code to speed up the 6502/6510 emulation, the speaker had it add debugging and inspection tools. These included a TCP interface for inspecting memory pages, CPU registers, and text or graphics memory, along with breakpoints based on simulated frames or the program counter. The agent could then run a graphics demo, inspect the emulator’s state when it failed, and iterate on its changes. Without these tools, it initially claimed the work was done even though the demo did not work. Adding scattered print statements was a poor substitute for proper instrumentation.

• Dump1090 and ADS-B decoding: While exploring neural-network-based phase correction for signals received with RTL-SDR devices, the speaker first established comparison baselines, including existing phase-correction methods. He also arranged for the program to dump packets it failed to decode, allowing the agent to inspect whether a packet contained an error or whether phase correction was inadequate. This made progress measurable and helped the agent refine its approach.

• Frontend work: For visual tasks, agents should be able to use a browser such as Chrome to take screenshots and inspect the layout. The speaker also recommends finding a simplified way to represent the layout so the agent can assess it, rather than relying only on written instructions.

The general recommendation is to create observable systems and objective feedback loops: expose relevant internal state, capture failures in a form the agent can analyze, and provide ways to assess whether changes actually worked. Tests are valuable, but for complicated problems they may not provide all the insight that direct inspection can.

Main Speaker and Source

Salvatore Sanfilippo speaks from his experience with coding agents, including the Commodore 64 emulator and Dump1090 examples. Claude Code is the coding agent used in the examples.