The beginning
I wanted to make invisible context visible.
I use AI for coding, writing, planning, and problem solving. Early on, I noticed that long conversations became harder to manage even when the latest message was short. The active context included far more than the words I had just typed.
That was the idea behind the first Tangerine. It counted tokens, estimated how a session could grow, and tried to help people make prompts more efficient. The point was not perfect arithmetic. It was helping people understand what the system was carrying.
The first version did what I intended. It gave people a simple way to see an invisible part of AI use and understand why a conversation could become heavier over time.
What changed
AI work stopped looking like one long conversation.
People now use agents that can inspect files, call tools, run code, and carry a task through several steps. Products use APIs to bring AI into repeatable workflows. Teams can divide independent work across several agents and combine the results.
In that world, shaving a few words from a prompt is rarely the most important decision. The larger questions are about scope, access, state, evidence, and how much autonomy the system should have.
What stayed
Context still shapes the quality of the work.
Tokens still matter because they carry instructions, source material, tool results, history, and the final response. But efficiency should mean keeping the useful context and removing the clutter, not simply forcing every request to be shorter.
The durable lesson is the same: understand what the system is carrying, give it a clear job, and require evidence that the work is complete.
Where it is now
A practical guide to choosing how AI should work.
Tangerine now explains the difference between a direct response, a single agent, an API workflow, and multiple agents. It also shows how instructions, tools, files, history, and coordination add complexity to the working context.
It does not pretend that a simple rule can classify every task. The goal is to give people a useful mental model so they can make that decision themselves.