On August 26, 2026, Dolores Research released WAGMI Bench, an open-source benchmark for evaluating AI trading agents in Bitcoin perpetual futures across 13 historical periods, with a design that elevates survival and participation over raw profit. The project is Apache-2.0 licensed and invites public use and modification, according to Pluang.
The shift is subtle but important: agents are judged by whether they stay alive and engaged through many regimes, not just how much they made in a lucky window. That framing aims to reduce cherry-picking and force apples-to-apples comparisons. In a market where liquidation risk can erase months of gains in minutes, this is overdue.
What WAGMI Bench changes in AI crypto testing
Profit-centric backtests incentivize fitting models to a handful of favorable periods. WAGMI Bench standardizes the test across 13 fixed windows, making it harder to game the setup and easier for others to reproduce results. That reproducibility—run the same code, get the same outcomes—is how machine learning moved forward in other domains. Think of how MLPerf did it for training and inference scores in mainstream AI; shared rules raised the bar for fair comparison (MLCommons).
Overfitting is the oldest trick in quant. Researchers have shown how easy it is to dial a strategy until it shines in-sample, only to fail out-of-sample. A widely cited paper on the probability of backtest overfitting lays out how selection and tuning can create mirages. By locking in multiple historical spans—and weighting survival over peak return—WAGMI Bench pressures builders to ship agents that generalize, not just memorize.
Per Pluang, the benchmark also emphasizes engagement, a guardrail against agents that simply sit flat to avoid mistakes. That matters for market utility. A bot that never trades can look safe on paper, but it’s not useful liquidity.
Why survival-first metrics beat pure PnL
Perpetual futures come with funding payments, leverage, and the ever-present risk of forced liquidation. In this arena, a strategy’s first job is simple: don’t die. A survival-first objective encourages capital preservation, limits tail exposure, and forces drawdown awareness. For readers new to the product, here’s a quick primer on perpetual futures and why funding and leverage make ruin a real risk.
Pure PnL metrics can reward lottery tickets. One massive win in a bull stretch can mask fragility that shows up the moment volatility flips. A survival-oriented score, by contrast, tests whether an agent can handle trend, chop, and shock. That’s the promise behind WAGMI Bench: filter for resilience across regimes before anyone turns a bot loose with leverage.
There’s also the public-good angle. If builders gravitate to strategies that avoid sudden liquidation, venues see steadier participation and fewer cascade events. That’s better for the market microstructure and for users on the other side of those trades.
How builders can use WAGMI Bench today
Because it’s released under the Apache-2.0 license, teams can integrate the framework into their CI pipelines and share results without legal friction. A practical flow:
- Run your agent across the 13 fixed historical periods before any live experiment. Treat that as a go/no-go gate, not a demo.
- Set explicit limits on leverage and position size that reflect the benchmark’s survival focus. Optimize for staying power, not just upside.
- Publish seeds, configs, and unfiltered logs so others can reproduce every trade your agent took. If someone else can’t rerun your test, it doesn’t count.
- Stress scenarios with fee and slippage assumptions that err on the high side. Backtests that ignore frictions aren’t tests; they’re fiction.
For research teams, WAGMI Bench is also a standard to teach against. Students and new hires learn quickly what matters when “don’t get liquidated” sits at the top of the rubric. For exchanges and prime brokers, accepting results formatted to a shared template could streamline risk reviews.
Token design and governance: reading $DOLORES
Alongside the framework, Dolores launched a community token, $DOLORES, connected to the Virtuals Protocol on Robinhood Chain. Pluang reports that it’s designed for research coordination rather than governance. That distinction matters. Governance tokens can create complex expectations about control and rights, while a coordination token signals funding and participation without conferring the keys to the protocol itself.
It also hints at a model many open-source AI projects are testing: separate the artifact (code and datasets) from decision rights, then use a token to nudge collaboration and maintenance. If it channels attention and grants toward the benchmark—documentation, new periods, adapters to different venues—it could speed adoption. If it distracts from the core mission, it won’t.
What success would look like for this open-source benchmark
Benchmarks rise when the rules are clear, the tasks map to the real world, and the results are easy to compare. WAGMI Bench has a head start on the first two: Apache licensing, public code, and a test bed tied to BTC perps through diverse history, per Pluang’s summary. The comparison layer will depend on whether leading teams publish their runs, and whether venues or funds accept those runs as part of diligence.
If that happens, expect two second-order effects. First, fewer “hero” equity curves based on one cherry-picked regime. Second, more work on agent safety—position sizing, circuit breakers, and adaptive hedging—because that’s how you score under a survival-first rubric.
There’s a broader lesson here too. Open-source benchmarks shape behavior. In NLP and vision, the community over-optimized to leaderboards at times, but it also learned to spot shortcuts and stale tests. Crypto markets change fast, so the caretakers of WAGMI Bench will need to keep adding periods, especially stress episodes, and make it easy to audit any change log.
Pluang also frames the launch against improving infrastructure and a warming regulatory climate for AI-driven trading. If compliance teams can point to a common, reproducible test that penalizes recklessness, they’ll have one more concrete artifact to review. That won’t settle policy debates, but it moves them from vibes to evidence.
The work now is adoption. If builders, exchanges, and research groups align on shared tests and transparent reporting, WAGMI Bench could become the default screening tool for crypto-trading agents—less sizzle, more staying power.