Ways to interact with LLMs

Rusty Turkek described an interesting way to use LLMs to "unclog" his mind from what I take are too many ideas: Let an LLM do the ruminating. (see thread)

That might be more of a niche application, but Paul Graham asked a follow-up question, and that led to the catchy phrase of "using AI in batch mode". That felt vaguely familiar and I immediately wondered whether there are ways between "interactive mode" and "batch mode" that better fit what I am doing and so I replied with the two ways I use LLMs. I have since come up with this list:

• realtime mode - when you use voice mode and the exchange is by sentences (or less, if you interrupt).
• interactive mode - chatting back and forth, using short sentences like in a messenger. This seems to be the most common mode. Often used with a fast/instant model.
• block operation - this is when LLM outputs are long, and you respond by addressing multiple points in one lengthy reply too. This is the letter exchange I mentioned in the tweet. I use this typically with high reasoning settings. I use this often at the end of a conversation to generate a paper or experiment to convert into an artifact I can export and reuse.
• orchestration - when you have a complex task that requires frequent oversight because the agents do not understand the domain sufficiently to let them run unattended. Common in coding in general. You may want to tell your agent to ask questions and use e.g. Karpathy's CLAUDE.md.
• batch mode - when you specify a complete task that you can let the model work on unattended and only need to see the result. Rusty Rurek above used this. Depending on the size of the task, this may require a Pro model or scaffolding with /goal, e.g., in an IDE like Cursor, Codex, Claude Code. Deep Research is like this naturally.
  • plan mode - before starting batch mode, the actual command may be developed with a "plan mode" where the agent cannot change anything except a plan that may be built incrementally, e.g., using the above modes.
• swarm mode - when the task requires a breakdown into multiple subtasks that can be (or have to be) worked on independently. Agents in IDEs start this easily on repetitive tasks even unprompted because they have tools that allow the spawning and monitoring. I use this, for example, on research tasks that require blinding, i.e., where the agent scoring a record should not know the purpose of the scoring^[1].
• infrastructure mode - this is when you build an infrastructure of multiple agents, some of which may run continuously or on schedule. At AE Studio, we did that a lot.

I'm sure you can up with more such "modes".

What is interesting is that interacting with coding agents changes the flow of coding. As Gordon Seidoh Worley observed in Every Engineer a Manager: You have more interruptions or at least change topics more often. As an oldtimer, this feels a bit like working on another program while waiting for the compiler. But it is not really the same schedule of a maker, where your calendar is filled with hourly slots as discussed by Paul Graham in Maker's Schedule, Manager's Schedule. You have much more flexibility in managing your agents. And it requires less social skills.

1. ^{^}

   This is surprisingly tricky to get right because the tool calls typically can't isolate the sub-agents sufficiently, and you have to work with cloned workspaces that are cleaned before the task. Or you need to build something more specialized yourself.