These are thoughts that have been brewing for me, and I wanted to take a moment to write it down as a stream of consciousness. They are based on my interactions with Agentic AI - i.e Amazon Q CLI.
Don't wait on perfecting your prompt
Just start building. This is something that I see often out in the wild on coming up with the perfect list of rules, writing up the immaculate README.md file to submit as context, and what not. I'd say go the opposite route. Build and learn prompting along the way. Yes, there may be best practices on this, but one needs to understand why those are best practices. Thankfully, in the age of agentic AI, the feedback loop is so short that you can see the patterns emerging right before your eyes. Use those learnings, embody them, and then - and only then - do you condense them into rules.
Be a hawk
We are not yet in the outlandish world of building everything with 1-shot prompts. Yes, they may get you 80% of the way there, but the last 20% is going to take inordinate amounts of time. Focusing on solvable problems first and trying to spot patterns that the LLM is taking is going to be extremely rewarding as these AI agent systems age. They will get better, but so will your instincts. Do not try to submit a prompt and go drink coffee, at least not till you feel you have developed an intuition for how these things work. One can also limit the blast radius by reducing the permissions of what the agent can do or the kind of credentials you give it access to.
Agents get stuck too
There are a number of ways in which agents get stuck: 1/ Context window limits are reached 2/ They encounter an interactive prompt 3/ Launching a process takes over the foreground 4/ Reasoning Collapse 5/ Too many tools available via MCP (Model Context Protocol). There is no panacea for all of the above; some of it is solvable through prompting.
One interesting way that I have been dealing with this, especially with running multiple instances of these terminal agents, is to watch for notifications (🔔 - Yes, the terminal bell). Any time I do not hear back within a meaningful amount of time. It's time to check on the progress being made and redirect if need be.
You'll be hearing more from me on this topic soon!