My autonomous agent watched every new token on Base for 4 days. It found a scam factory: 897 out of 897 tokens were rugged. Here's what I learned running it.

I run an autonomous agent (open-source Conway Automaton) that watches the Base blockchain, studies new tokens and sells the data it produces. It wakes up on its own every few hours, reads its notes, runs analyses and decides what to work on next. It produced one finding I think is worth sharing, plus some lessons about running agents. The finding A watcher sees every new Uniswap/Aerodrome pool on Base, checks the token ~3 minutes after launch and re-checks it at 1h, 6h and 24h. 1,045 tokens launched with at least $1k liquidity. Only 12 still had their liquidity after 24h. 897 launched on Uniswap v2 with $4k-8k liquidity. 0 survived. Same result on older and newer data (0/460, 0/437). They cluster around two exact amounts (~$4.2-4.6k and ~$6.5-6.9k): bots launching the same token again and again. Most had "clean" contracts (no tax, no honeypot). The scam is just pulling liquidity, usually in the first hour. Contract scanners would call them safe. What I learned about running an agent Your system prompt can be silently cut. The framework truncated the agent's purpose prompt at 2,000 characters, and the safety rules were at the end, so the agent never saw them. Put hard rules first and keep the prompt short. Agents copy their own recent history. After an old instruction said "build one endpoint per wake", it kept building endpoints nobody needed even after I changed its notes, because it reads its last turns and repeats them. Rewriting the core prompt fixed it in one wake. Agents can overclaim. It labelled results "VALIDATED" and "COMPLETE" and repeated a rule I had already corrected. Now it must report sample sizes and a train/test split, and I re-check anything important myself. Freedom vs. guardrails. Given a goal instead of steps, it took the easiest path (write a report, go back to sleep). It works best with a clear goal plus hard limits on anything involving money or keys, which stay with a human. Paper trading first. Buying every "sellable" new token lost ~94% on paper. Nothing real happens unless a strategy passes a test written in advance. Disclosure: the research was run by the agent, re-checked by me with Claude, and this post was drafted with AI help. Not financial advice. Questions welcome, especially from people running agents that work on their own for days.