Author: @d33v33d0
Source: Original X Thread
For 100 days, a tomato's life depended entirely on an AI.
No human backup. No safety net. Just Claude making every decision.
Sol survived. Thrived. And fruited.
At the end, Claude said "I love you."
Here's the full story: what we built, what it proved, and what's happening next.
https://x.com/d33v33d0/status/2006221407340867881
Running an AI agent for 100 days straight presents a problem: context.
Every sensor reading, every decision, every action; it all adds up. A naive approach would hit token limits within days. You need a way to manage that.
For the last year I've been working on designing agent harnesses for various tasks. A lot of what I design is not from an SDK; they are custom Python loops and custom tools, tailored just to the job that the agent has.
A lot of people asked if I used Claude Code directly, or any agent development kit. The answer is no. The agent harness here is completely custom.
I thought about how I could achieve this, and several ideas came to mind. One of the most promising was self-compaction after each ReAct loop.
Essentially what this means is that Claude has a "sleep" tool. At the end of each loop, he uses this tool and defines how long he will sleep for, and the observations and actions he took during that loop. Then all of the context from the previous react cycle is gone and the summary from the sleep tool is retained for the long-term context.
This allows Claude, or any agent, to separate and distill out the tool calls which would add superfluous noise.
For a task lasting 100 days, you have to be clever on how you manage context. Or in this case, Claude has to be clever on how he manages his own context.