Agentra docs
Gemini
Gemini reads the crowded tape and helps the runtime keep names apart.
Place on the loop
Gemini’s place is the crowded tape. The trench is often several names at once, with symbols that look alike and contracts that do not. A runtime that only sees the newest pair can still use a second read when the tape is noisy: which names are actually distinct, which mention is a repeat of the pair it already bought, which burst of attention is the same mint said three ways.
That place sits beside GMGN, not instead of it. GMGN is the tape. Gemini is a reader OpenClaw can use while looking at that tape. The cycle still ends as a Solana buy and sell. Gemini does not become the market.
The read
The useful read is separation. Given a handful of symbols and mints, which ones are the same route and which ones are neighbors. The activity list depends on that separation, because a later page of buys and sells is only trustworthy if two similar tickers were not fused into one story.
The read is also about noise. A trench name can arrive with extra words, line breaks, or a symbol too long for a console column. The feed already cleans a symbol down to a short printable name. Gemini’s job is the judgment around that string: whether the name in front of the runtime is the name the tape is actually trading, or a caption riding beside it.
Shared with the runtime
What Gemini shares with OpenClaw is a clearer set of candidates, not a signed transaction. The runtime still applies the mode, still takes size from the vault, and still prefers the newest Pump.fun pair when it is time to buy. A clearer candidate list helps that choice. It does not skip it.
Because the loop is shared, Gemini also sits next to OpenAI and Claude. OpenAI is closer to a prompt you typed. Claude is closer to the recent book. Gemini is closer to the live tape. If you remember the three as tape, prompt, and memory, you will not expect one of them to do the other two jobs.
Boundary
Gemini does not custody the vault and does not set the win rate. A good read can still be followed by a loss. Losses are part of the desk, in the 0.5 to 1.1 SOL band, and they are recorded like wins. If a page about a model makes a loss sound like a failure of the model, the page is wrong. The close is a property of the route. The model’s page ends at the read.