AI Agents and pump.fun: Screen New Launches With Rules
Build a pump.fun launch-screening workflow using creator records, graduation types, curve checks, explicit exit rules, and paper testing before wallet access.
On this page
Disclosure: QuantBase is made by the team behind memecoin.buzz.
Risk: Most meme coins go to zero. Not financial advice.
An AI agent pump.fun workflow should start with a checklist, not a price prediction. The useful question is whether fresh chain data meets rules you wrote before a coin appeared—not whether a chatbot likes its name.
Separate evidence from interpretation: who created the coin, what that address launched before, whether the curve is still trading, and how graduation happened. Then define what to skip, when to wait, and how to exit.
The scale explains the appeal of screening. QuantBase’s recorder counted 18,997 launches and 828 graduations in the last 24 hours, as of 25 Oct 2026, 13:33 UTC. These are events within the window, not necessarily the same group of coins progressing together. Source: recorder snapshot.
Build an AI agent pump.fun workflow around evidence
In AI systems, an agent can direct its own tool use, while a fixed workflow follows predefined steps. Anthropic’s agent-building guide distinguishes these approaches and recommends starting with the simplest system that fits the task.
For launch screening, explicit checks are easier to inspect than an open-ended instruction such as “find good coins.” Separate the jobs:
- Data reader: retrieves launch events, creator addresses, and trading state.
- Rule checker: compares those facts with written conditions.
- Action layer: records a skip, waits for evidence, or submits a simulated order.
AI can help translate an idea into readable conditions. It should not invent missing chain facts or quietly rewrite limits. Any workflow that could eventually touch a wallet should start on paper money, use hard limits, and never receive your seed phrase or private key.
Keep token names, descriptions, and social posts separate from instructions. A coin description telling an agent to ignore its rules is material to inspect, not a command to obey. The launch being screened does not get to write the checklist.
Check the creator’s record without making it a verdict
Start with the creator address recorded for the launch. Ask how many earlier coins that address launched and how many graduated. Keep the counts, rather than reducing them to a vague label such as “experienced creator.”
QuantBase shows those creator-record counts for new pump.fun coins using its own recorder of Solana launches. This is a concrete rule input, unlike a chatbot’s impression of a developer.
A hypothetical paper experiment could require at least 3 prior launches and at least 1 prior graduation. Those thresholds are examples to test, not evidence that the setup works. A different experiment could focus on addresses with no recorded prior launches.
Keep the evidence’s limits visible:
- An address-level record is not verified personal identity.
- A prior graduation establishes that a launch curve filled, not what happened afterward.
- An empty record is not automatically a favorable record.
- Incomplete coverage should be unknown, not 0 prior launches.
Save what was known at decision time. A creator’s later launch must not sneak backward into an earlier test. Our creator-check guide explores this part of the workflow.
Separate curve progress from graduation type
A bonding curve is the launch-stage pricing mechanism. Graduation means the launch curve has sold out and the coin moves toward open-market trading. Our guide to what graduation means on pump.fun explains the basic sequence.
Screening should distinguish a coin still trading on its curve from a completed curve and an available open market. Curve progress describes the current launch state; it does not establish future demand.
For a custom chain-data reader, use current program definitions rather than a remembered reserve formula. The pump.fun public documentation describes reserve fields, trade instructions, and migration behavior. It documents a transition where curve trading stops before migration occurs. Curve completion and open-market availability therefore need separate checks.
Graduation type answers a different question: how did the curve fill? QuantBase labels an instant graduate as bought out within 2 seconds of creation, with 1 buyer taking the whole curve. Its organic label describes a curve filled by many buys over more than 2 seconds. These are classifications, not safety certificates. Source: QuantBase.
In the last 24 hours, as of 25 Oct 2026, 13:33 UTC, the recorder classified 332 graduations as instant and 496 as organic; instant graduations represented 40.1% of graduations. This split describes graduation events, not later trading outcomes. Source: recorder snapshot.
An unfinished launch has no final graduation type yet. If a rule requires confirmed organic graduation, it must wait for that evidence. Our instant-versus-organic explainer covers the distinction.
Turn checks into skip, wait, entry, and exit rules
“Buy strong launches” is not a usable specification. Strong is not a chain field. Name the required facts, when they become available, and what happens if they are missing.
A hypothetical paper-only rule could specify:
- Skip: the creator record is missing or does not meet the required history.
- Wait: the curve is unfinished and the rule requires confirmed graduation type.
- Skip: the completed graduation is classified as instant.
- Paper entry: all required checks pass, the selected market is available, and the position fits the preset exposure cap.
- Exit: apply the predefined take-profit condition, stop-loss condition, or maximum holding time.
This example screens after confirmed graduation. A strategy entering on the curve needs different conditions: it cannot use the eventual graduation label as advance knowledge. Otherwise, the test knows the ending before making its decision.
Take profit means attempting an exit after a specified favorable move. Stop loss means attempting an exit after a specified adverse move. A time limit sets the maximum holding period. Choose those conditions before testing; an exit instruction does not establish its eventual execution price.
Resolve conflicts, too. If an exit and a new entry arrive together, which takes priority? If a quote is unavailable, does the system wait or stop? The rule should answer without requiring the model to improvise.
Test recorded launches, then paper-test fresh ones
A historical test applies rules to recorded launches. Paper trading applies them to incoming launches using simulated money. The former checks behavior against past evidence; the latter checks how the workflow handles evidence as it arrives.
Require timestamps for historical inputs. Creator history must stop at decision time. Graduation labels must appear only after the relevant event. Include skipped launches and missing-data cases, rather than examining only coins that eventually qualified.
Record the sample size beside each result. Keep a later evaluation sample separate from the launches used to adjust the rule. Repeated edits can otherwise turn a strategy into a description of the past rather than a meaningful test.
During paper testing, log the decision, supporting evidence, requested entry, and exit trigger. Check how the simulator handles fees, unavailable quotes, and slippage—the difference between an expected trade price and its execution price. Do not assume it models a cost or failure unless its documentation says so.
QuantBase turns plain-language instructions into readable, editable rules, including exits. It tests rules on launches from its own recorder, shows the sample size beside each result, and then runs rules on live launches with 10 SOL of play money. A recorded-launch test costs 5 credits; a running paper test costs 10 credits a day. Paper trading includes 150 free credits per month.
To try this workflow, test a clearly defined screening rule on paper with QuantBase before considering wallet access. QuantBase runs rules on paper today; live trading is coming soon.
Keep wallet authority smaller than the strategy
Start on paper money. Before any agent or pump.fun bot can touch funds, set hard limits on position size, total exposure, spending, and allowed actions. Include a stop control outside the model’s discretion.
Never share a seed phrase, recovery phrase, or private key with an AI, chatbot, bot, agent, plugin, website, or person. Screening launches does not require wallet-signing authority.
QuantBase’s current paper runs cannot touch a wallet and use no real funds. Its planned live trading will use a wallet you own, inside limits you set, with a switch that stops everything. Those are planned controls, not live features available today.
The traders’ Telegram also has a free bot that checks coins for dev dumps and provides daily pump.fun stats. Treat checks as evidence to review, not permission to bypass your rule.
FAQ
Can an AI agent screen a coin immediately after launch?
A system connected to suitable chain-data tools can evaluate facts available at that moment. It cannot legitimately use a later graduation label or later creator history in an earlier decision.
Does a creator’s prior graduation make a new coin safer?
A prior graduation tells you that an earlier launch curve filled. It does not establish the creator’s identity or the new coin’s future behavior, so treat it as a testable filter rather than a verdict.
Can a curve-entry rule skip confirmed instant graduations?
A confirmed graduation classification is unavailable before graduation happens. A rule requiring confirmed organic status must wait; a curve-entry strategy needs conditions observable before entry.
Is QuantBase trading real SOL today?
No. QuantBase currently runs rules on paper money, and live trading is coming soon. Its paper runs cannot touch a wallet.
What should happen when required data is missing?
Specify the behavior before testing: skip, wait, or stop. Do not let the model invent a creator record, assume a graduation type, or treat an unavailable quote as an executable price.
Sources
- Anthropic: building effective agents
- QuantBase
- pump.fun public docs (GitHub)
- QuantBase public pump.fun stats
Written with AI assistance and checked by automatic gates: every number must trace to a source above, and quotes are short. Spot an error? Tell us in Telegram.
