From Discovery to Executable Evidence: An Agentic AI Architecture for Trading Newly Launched Tokens Primus Vekuh Research manuscript | 24 September 2026 Abstract Newly launched tokens create a difficult decision problem: the most visible price increases can occur before reliable information about liquidity, ownership concentration, sellability, and execution costs becomes available. We describe an implemented, multi-service research system that separates market discovery, short-horizon candidate assessment, pre-trade verification, paper execution, live execution, position supervision, and post-trade audit. Its central design principle is that a predicted rise is a different proposition from an executable and recoverable return. We formalize that distinction as an evidence-constrained decision process and illustrate it with records from Solana and Ethereum-compatible token experiments and adjacent foreign-exchange and liquid-crypto cohorts. A September 18, 2026 operational report records 23 Robinhood early-mover paper signals and 5,024 rejected-control observations but no realized closes in the active cohort; two older paper positions lacked fresh executable sell quotes. Forty Solana candidates were evaluated by each of two experimental controls, with none qualifying in that snapshot. Historical losses in other cohorts and incomplete fee and receipt evidence rule out a claim of demonstrated profitability. The research contribution is a falsifiable architecture and evaluation protocol: preserve point-in-time provenance, measure discovery-to-fill attrition, distinguish marks from realizable exits, and compare frozen policies prospectively under explicit costs. This case study shows why an agent can improve the quality of a trading decision by identifying what it does not yet know, even when doing so prevents an attractive retrospective trade. Keywords: agentic AI; meme coins; token discovery; algorithmic trading; execution; point-in-time evaluation; risk accounting; multi-agent systems. 1. Introduction A trader opening an application after a token has risen several thousand percent sees a simple story: an early purchase would have produced an extraordinary gain. At the moment of the proposed purchase, however, the trader would have had a different information set. The token might have been minutes old; a quote might have referred to another pool; a sell route might have been absent; a quoted price might have been too small to support the proposed order. A chart records one price path. A trading system must establish whether that path was discoverable, actionable, and reversible at the relevant time. We study this problem through the development of the Primus Trading Desk, a collection of services and experimental policies for discovery and evaluation of newly launched tokens on Solana and Ethereum-compatible venues, alongside comparison and operational reporting for other asset classes. The work grew from a practical question: why were rapidly rising tokens missed while screened candidates sometimes lost money? That question requires more than a faster predictor. It requires an account of every transition between seeing a token, deciding to buy it, acquiring it, observing its value, and selling it. The term agentic here refers to a system that observes an environment, calls specialized tools, maintains state across observations, makes conditional decisions, and revises its assessment as new evidence arrives. It does not imply that a language model directly controls a wallet or that the named strategy versions are trained neural networks. The available operational records establish specialized services, deterministic strategy identifiers, research comparisons, and evidence-aware reporting; they do not establish the internal training procedure of every model. This precision matters because research on language-model agents [1–3] does not by itself validate a particular live trading implementation. Our contributions are threefold. First, we specify an evidence-constrained architecture for newly launched tokens. Second, we offer a reproducible evaluation design that measures missed opportunities and false confidence under the same point-in-time data constraints. Third, we analyze operational snapshots, including negative and incomplete results, to show where the architecture currently fails to establish executable profitability. 2. Related work and research gap ReAct demonstrated a general pattern in which reasoning and tool use alternate [1]. FinMem introduced structured memory for financial decision making [2], while TradingAgents studied specialized analyst, trader, and risk roles [3]. These ideas motivate role separation, but a token market adds an unusually sharp gap between a textual or numeric recommendation and a feasible transaction. A plausible token identity, a fresh price, and a successful sell at the relevant size are separate facts. Empirical work on token ecosystems documents short token lifetimes, anomalous issuance, and rug-pull behavior [4]. Automated-market-maker execution also exposes traders to slippage and transaction-ordering effects [5]. These findings motivate checking creator behavior, liquidity, pool identity, and the economics of an exit. Bailey and colleagues show how selecting attractive historical strategies can inflate apparent performance [6]. Accordingly, our proposed test freezes policy versions before each forward window and does not transfer wins across versions. NIST’s AI Risk Management Framework supplies a broader rationale for traceability, monitoring, and explicit uncertainty [7]; it is not evidence that a strategy earns a return. The gap addressed here is operational: much agentic-finance research evaluates the quality of a decision or portfolio trajectory, while a new-token system must also measure the changing availability of evidence and the probability that the recommended round trip can actually be completed. We treat refusal, missing data, and late discovery as observable outcomes rather than silently deleting them from the denominator. 3. System and decision model 3.1 Observations, identity, and provenance At decision time t, a candidate observation is [o_t=(c,,p,s,t_s,t_r,x_t),] where c is contract or mint, χ the chain, p the pool or route, s the source, t_s the source timestamp, t_r the receipt timestamp, and x_t the measured features. Features may include price and volume changes over 5, 15, 30, 60, 120, and 240 minutes, depth, concentration, creator outflow, estimated sell impact, and route availability. A ticker is never sufficient identity: different contracts can share a symbol, while one contract can trade in multiple pools. A historical feature is admissible only if its source time and ingestion time precede the decision. This avoids a common form of hindsight leakage. The architecture has six logical stages: discovery gathers possible tokens; enrichment joins market and contract observations; policy evaluates a signal; an execution verifier requests size-specific buy and sell routes; a position supervisor monitors marked and executable exits; an audit process reconciles receipts and reports. A candidate can fail at any stage, with a structured reason and timestamp. The services reported under Solana early discovery and execution, Robinhood early movers and direct swaps, Coinbase research, and the consolidated reporting desk implement parts of this architecture. The available evidence is sufficient to analyze their reported behavior, but not to assert a complete independent source-code audit. 3.2 Separating a signal from an executable trade Let q be the proposed order size, B_t(q) the executable cost of buying that size, and S_{t+h}(q’) the executable proceeds from selling the acquired quantity at horizon h. Define [{t,h}(q)=S{t+h}(q’)-B_t(q)-F_{t,h}-G_{t,h},] where F is fee and tax cost not already embedded in the quotes and G is network and other settlement cost. Components must be itemized so spread or slippage is not charged twice. The gain shown on a chart, (q(P_{t+h}-P_t)), is not a substitute for (). If a fresh sell quote or the resulting quantity is missing, () is unknown, rather than zero or the last displayed market price. A buy decision is conditioned on policy eligibility, evidence freshness, route feasibility, and exposure. A simplified rule is [a_t={E[{t,h}o{t}]>, V_t(q)=1, R_t(q)L},] where V verifies the evidence and routes, R is proposed risk and L is the applicable limit. This expression describes the research design, not a verified transcription of every production strategy. Unavailable evidence should produce an explicit unknown or rejected for missing evidence status. Relaxing a confirmation threshold may increase early entries; it cannot create a sell quote that does not exist. 3.3 Agent roles and constrained adaptation An orchestrator can ask a discovery agent for new pools, a market agent for point-in-time momentum, a contract agent for concentration and control risks, an execution agent for size-specific routes, and a monitor for exit conditions. A research agent groups failures and proposes policy variants. These roles can be separate services or modules; we do not claim that each is an independently trained language model. The agentic feature is the conditional tool workflow and its recorded memory of decisions. Research proposals belong in a versioned challenger lane. Promotion requires a prespecified comparison on future observations, with the original candidate universe and rejected controls retained. A model must not select an attractive date range, rename it as a new version, and discard earlier losses. Human-defined limits on exposure and execution remain part of the experimental design even when analytic tasks are automated. 4. Materials and methods 4.1 Records and units of analysis We examined two user-provided operational artifacts: the September 8–9, 2026 trading review [8] and the September 18, 2026 hourly trading overview [9]. They are selected case-study records, not a complete machine-readable census of all events. Their units differ: candidate observations, rejected controls, signals, position opens, position closes, and broker-confirmed executions. Repeated quote updates are not independent trades. A single strategy’s historical result is not merged with another strategy or with an open position. For reproducibility in a future full study, the raw event table should contain: immutable event ID; chain, contract, pool, and strategy version; source and ingestion times; candidate features as known at decision time; decision and reason; quoted buy and sell amounts at proposed size; route identifiers; simulated or actual transaction identifiers; before-and-after balances; fees; subsequent marks; and terminal state. Data access and privacy controls may require releasing a redacted event table instead of wallet credentials or private infrastructure details. 4.2 Outcomes and comparisons The primary endpoint is net realized return per completed, independently identified round trip, with actual receipts and asset deltas for live positions and explicitly modeled executable costs for paper positions. Secondary endpoints are detection latency, signal-to-qualified conversion, qualification-to-fill conversion, time to first fresh sell quote, fraction of unpriced positions, worst plausible recovery under specified stress, and maximum adverse excursion. We report win rate only alongside the distribution of gains and losses; a high win rate does not imply positive expectancy. To test whether role separation and evidence checks help, a prospective experiment would compare: (A) a fixed momentum-only policy; (B) the same policy plus point-in-time liquidity, identity, and route checks; and (C) the full orchestration and exit supervision. Each arm must see the same candidate stream and proposed notional, using common timestamps. A separate ablation removes each evidence source without changing other thresholds. A blocked candidate must remain in the denominator and be revisited at fixed horizons with historically available executable quotes, never a retrospectively selected peak. Bootstrap intervals should resample independent tokens or market episodes, not repeated observations of the same pool. Policy selection belongs to a development window; final evaluation belongs to a later untouched window. 4.3 Evidence status We distinguish observed live broker-confirmed closes, live reported activity requiring receipt reconciliation, paper simulated fills, shadow paper counterfactual policies, open marked unrealized positions, and unpriced positions without a usable current exit quote. A stress loss at zero recovery is a scenario, not an executed sale. These labels are essential to interpreting the results below. 5. Observations from the case study 5.1 The early-token funnel The September 18 overview labels the active Robinhood early-mover cohort as 23 signals and 5,024 rejected-control observations, with zero closes [9]. These counts describe the reporting state of that cohort, not 5,047 independent tokens and not proof that the accepted signals were profitable. Its horizon-specific observations cover 15-, 30-, 60-, 120-, and 240-minute intervals; their counted controls and signals are overlapping observations rather than independent trials. The same report lists two older Robinhood paper positions, CUP and WORM, without fresh executable sell quotes. It excludes them from marked P&L instead of treating their last observed prices as cash recovery. The Solana Early Control and experimental Pump.fun expected-value control each evaluated 40 candidate observations and qualified none in the September 18 snapshot [9]. The report flags missing creator-selling information for 40 observations and unavailable paper creator-concentration information for 35; it also records failed discovery-to-entry refresh and handoff checks. Those are concrete sources of attrition. They do not prove whether the rejected assets would have been profitable to trade at executable size. 5.2 Accounting changes the story The same September 18 report displays +$25.3547 of reported open paper P&L across nine priced positions, with two more unpriced, and zero live and paper closes in that hour [9]. The open amount is not realized profit. The report labels four marked assets as having a mixed accounting basis; for example, BTC has +$1.0593 of raw price movement but +$0.6951 reported open P&L, an implied adjustment of -$0.3642 that is not a verified fee breakdown. Thus even a positive displayed mark can be insufficient evidence of net recoverable profit. Historical cohort summaries in that report include 45 observed live Forex closes totaling -$0.5755, five observed live Coinbase closes totaling -1.4227 USDC, and twelve Solana unified paper closes totaling -$1.8472 [9]. These figures are denominated in different units and account for different periods and policies; adding them would be misleading. The Robinhood direct-swap line states that a verified fill cannot be inferred from wallet funding or service readiness. These observations directly contradict a blanket claim that the deployed system has demonstrated a profitable edge. The earlier September 8–9 review gives a small, instructive path comparison [8]. Two $2 Robinhood paper positions, BITCAT and GARY, closed at -$0.512199 and +$0.182922, respectively, a combined -$0.329277. BITCAT subsequently displayed a positive model-path value at a later horizon, while GARY fell after its protected close. Neither retrospectively favorable exit point was a guaranteed executable fill or a prospective rule. The opposite paths show why both rapid exits and patient holds can look persuasive after the fact. 5.3 What these observations establish The evidence establishes an operating research workflow with candidate screening, versioned cohorts, paper histories, rejection reasons, and more careful separation of realized, unrealized, and unpriced outcomes. It identifies specific missing inputs and execution-verification problems. It does not establish improved returns from agentic orchestration, a 70% win rate, a reliable ability to capture thousand-percent runners, or successful Robinhood live round trips. The available documents are snapshots and reviews, not a complete independently audited dataset. 6. Discussion The practical tension in early-token trading is that more verification may delay an entry while less verification may admit an asset whose displayed value cannot be recovered. This is a measurable trade-off, not a reason to assume either maximum caution or unconditional entry is optimal. A useful experiment tracks the first time each evidence item became available, then measures whether a signal still had positive net executable expectancy at that time. It also records missed gains among rejected candidates and losses avoided by those rejections. The proposed architecture changes the learning target. Instead of optimizing a model solely to predict a high future chart price, the learner estimates the joint likelihood of timely discovery, admissible evidence, buy execution, valid exit, and net return. If these conditional stages are (D,V,B,X), then the event of a profitable completed trade depends on their intersection, not merely on a favorable price movement. This makes operational failures such as stale feeds or absent sell routes first-class research outcomes. There are threats to validity. The selected reports may omit earlier scans and contain overlapping summaries; strategy versions and deployment rules changed during observation; paper fills may not match live execution; and the same token may appear in multiple horizon counts. A screenshot of a later price spike selects a survivor after the outcome is known. Small completed-trade samples produce unstable estimates, particularly when payoff distributions have heavy tails. The paper therefore reports descriptive case evidence and an evaluation protocol rather than a causal estimate of the architecture’s effect. 7. Next study and conclusion The next study should freeze the candidate universe, policy versions, order sizes, source-time rules, and accounting definitions before collecting a forward sample. It should retain every rejected token, reconcile every live round trip to receipts and balance changes, and publish both the profitable and losing paths. A preregistered comparison of momentum-only and evidence-constrained policies can then answer whether the additional agents improve net executable return, and at what cost in missed early entries. The present result is narrower and useful: the implemented system makes it possible to ask better questions about missed meme-coin moves. Its operational records already expose where apparent opportunities become uncertain: at identity joins, stale observations, candidate handoff, route verification, and valuation of open positions. Until prospective, cost-complete round trips demonstrate otherwise, the scientific claim is an architecture for auditable decision making, not a profitable trading strategy. Data availability and disclosures The operational artifacts [8, 9] are user-provided reports. A public replication package was not available for this manuscript. Numerical examples are transcribed from those reports and should be checked against exported event and transaction logs before journal submission. The author should confirm affiliation, coauthors, financial interests, any ethics requirements, and permission to publish platform-specific or wallet-level details. No external funding or conflict-of-interest declaration is asserted here. References 1. Yao, S., et al. (2023). ReAct: Synergizing reasoning and acting in language models. International Conference on Learning Representations. https://arxiv.org/abs/2210.03629 2. Yu, Y., et al. (2024). FinMem: A performance-enhanced LLM trading agent with layered memory and character design. Proceedings of the AAAI Symposium Series. https://ojs.aaai.org/index.php/AAAI-SS/article/view/31290 3. Xiao, Y., Sun, E., Luo, D., & Wang, W. (2024). TradingAgents: Multi-agents LLM financial trading framework. arXiv:2412.20138. https://arxiv.org/abs/2412.20138 4. Cernera, F., La Morgia, M., Mei, A., & Sassi, F. (2022). Token spammers, rug pulls, and sniperbots: An analysis of the ecosystem of tokens in Ethereum and in the Binance Smart Chain. arXiv:2206.08202. https://arxiv.org/abs/2206.08202 5. Heimbach, L., & Wattenhofer, R. (2022). Eliminating sandwich attacks with the help of game theory. arXiv:2202.03762. https://arxiv.org/abs/2202.03762 6. Bailey, D. H., Borwein, J. M., López de Prado, M., & Zhu, J. (2017). The probability of backtest overfitting. Journal of Computational Finance, 20(4). https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2326253 7. National Institute of Standards and Technology. (2023). Artificial intelligence risk management framework: AI RMF 1.0. https://doi.org/10.6028/NIST.AI.100-1 8. Primus Trading Desk. (2026, September 9). Trading review, September 8–9, 2026 [User-provided operational review; unpublished]. 9. Primus Trading Desk. (2026, September 18). Hourly trading overview, 9:38 PM EDT [User-provided operational report; unpublished].