automated trading teams

Abstract: LLM trading agents are evaluated almost exclusively in backtests and autonomous benchmark harnesses. Deployment, where actions are irreversible, raises a different question: who verifies the agent's actions, and which failures can each verifier structurally catch? We report on AllPath, a personal LLM trading agent deployed against a live paper-brokerage account. The system watches positions through a deterministic rules engine, and each night it reviews its own day and proposes revisions to its own trading strategies. Verification is split across three layers. Code verifies what must not depend on judgment: deterministic gates enforce risk limits and block any strategy revision that fails an integrity check. The agent verifies itself within fixed bounds: its nightly review has an iteration cap and no tools that can act. A human verifies judgment, through approval screens that show the current state of each proposal. We also compared two verification policies for the agent's self-revisions on the live system: two and a half weeks with every self-revision human-gated, then five weeks with reflection revisions auto-applied behind the same code gates, while the operator's chat channel stayed human-gated throughout. Each layer caught a different kind of failure. The code gates rejected no proposal; their role was keeping revisions consistent. The human rejected two proposals that passed every code check, both over a disagreement about the market, and later removed 11 of the 19 rules the agent had added on its own, at a median lag of eight days. The most instructive production incidents, such as strategies silently left unmonitored, passed every gate; the operator found them by hand, and the fix each time was an interface change that makes the condition visible. We contribute the layered architecture, the two-regime account, and a production incident taxonomy, none observable in a backtest.

Abstract: We study when logged contextual-bandit data can certify an upper tail quantile for a policy selected from the same sample. We construct a PAC-Bayes Bernstein lower-confidence envelope for the target cumulative distribution function (CDF), simultaneous over a predeclared threshold grid and all data-dependent posteriors. The resulting procedure, TailCert, uses the envelope's first crossing of level p to return an atom-safe upper confidence bound on the selected posterior's pooled p-quantile (or inconclusive). It can also be used to select a policy under a certified tail constraint. The concentration ingredients are standard; the main theoretical contribution is a matched finite-class certification boundary comparing full logs with their joint weighted CDF indicators at a fixed threshold. On an explicit rare-action class, the required sample size grows linearly with the overlap factor W, logarithmically with policy-search and confidence complexity, and quadratically with the inverse CDF safety margin; the upper and lower bounds agree up to constants in the stated regime. On this class, the CDF reduction hides which actions exceeded the budget; retaining action counts restores the full-log rate. Simulations verify the predicted search-complexity shift, winner's-curse failure of naive reuse, and the power cost of splitting.

Abstract: Agents are deployed in applications from trip planners and stock trading to IT incident triage. In most cases, LLM agents work autonomously with minimal rule-based safeguarding, leading to cost and safety issues from irreversible actions. Recent works resolve this either by using a safeguard agent to monitor behavior or evaluating logs post-hoc. The first adds cost and latency to every step; the second delivers its verdict after the run, when tokens are burned and damage is done. To overcome this, we propose OnTrack, a streaming monitoring mechanism that compares an agent's steps and dependencies against recorded successful runs to alert users or block the agent in about a millisecond per step. We study this problem in three regimes of decreasing access: full reference access (historical runs and tool schemas), intermediate access (only tool schemas), and no prior knowledge (only step logs as generated). Expectation of OnTrack's monitoring capabilities reduces as data access drops, ranging from plan violation detection to identifying loops, stalls, and repeated tool calls. Finally, we evaluate OnTrack using SWE-bench trajectories. Based on the first 8 steps, our method ranks failing trajectories below succeeding ones better than content similarity approaches (+0.057 AUROC). With an abort policy, we save about 18% of compute that would be burned on failing runs, where 83% of interrupted runs were actually heading to failure (5 out of 6 aborts were correct).

Abstract: Online learning is an attractive framework in many domains because it permits well-defined learning even when data are dependent or chosen adversarially. This generality, however, comes at a steep price, introducing significant statistical and computational barriers. Recently, smoothed online learning has emerged as a promising framework that interpolates between the fully adversarial and fully stochastic settings by assuming that the conditional law of each covariate has density at most $1/σ$ with respect to some fixed base measure $μ$, and it is known to match the statistical and computational guarantees of classical learning while still allowing for much of the flexibility of online learning. However, existing oracle-efficient algorithms require either (i) sampling access to the base measure $μ$ or (ii) labels that are perfectly predicted by a fixed hypothesis. Both assumptions limit the applicability of these algorithms, in contrast to statistical learning, where empirical risk minimization (ERM) learns efficiently in the agnostic setting without any knowledge of the data distribution. We show that neither assumption is necessary, giving the first oracle-efficient algorithm that achieves sublinear regret in the agnostic setting without knowledge of $μ$. Our algorithm, based on Gaussian Follow-The-Perturbed-Leader, is parameter-free: it requires no knowledge of $μ$, the smoothing parameter $σ$, or the horizon $T$, and it achieves regret $\widetilde O(d\sqrt{T/σ})$ for binary classes of VC dimension $d$ with a single call to an ERM oracle per round, which is optimal up to a $\sqrt{d}$ factor. En route to establishing the regret bound, we introduce several new techniques that may be of independent interest.

Abstract: Diffusion policies represent multimodal action distributions, but an advantage-weighted update does not specify how to sample from the resulting target distribution. We propose Schrödinger--Föllmer Actor--Critic (SFAC), an offline-to-online reinforcement learning (RL) method for Kullback--Leibler (KL)-regularized policy improvement through conditional diffusion. A minimax Bellman critic estimates the advantage function, which defines an exponentially tilted target policy. A Doob $h$-transform expresses this update as a correction to the reference diffusion drift. We derive a posterior-mean representation of the correction and estimate it using paired self-normalized importance sampling (SNIS). Supervised regression on these drift targets updates the neural actor without critic action gradients. In the small-update regime, the KL-regularized update follows the natural policy-gradient direction, and the Doob correction represents the same local change in the space of diffusion drifts. Under suitable conditions, we derive finite-sample bounds that separate the effects of critic estimation, neural drift regression, finite-sample SNIS, diffusion discretization, and inherited actor error on expected average policy suboptimality. Synthetic experiments assess the accuracy of approximation to prescribed advantage-tilted targets and sensitivity to sampling budgets. On six offline-to-online continuous-control tasks, a reference-anchored implementation achieves higher final-window returns than those of its corresponding offline initialization.

Abstract: Adversarial multi-objective bandits hold the potential to help us optimize choices (arms) whose reward is a multidimensional vector chosen by an adversary and whose performance is measured by Pareto regret. We define loss as one minus reward and measure the easiness of a coordinate by the smallest cumulative loss of the arms on it, and call the coordinate easier when this quantity is smaller. Existing work suggests that in theory an easier coordinate may reduce Pareto regret. However, in practice, one may not know which coordinate is easier. On the negative side, we show that this lack of information eliminates the possibility: a smaller cumulative loss does not improve the worst-case order of Pareto regret. Precisely, let \(L_d\) be the smallest cumulative loss along coordinate $d$ over $T$ rounds. For \(K\ge4\) arms, \(T\ge6\) rounds, and at least 2 coordinates, we prove that the minimax expected Pareto regret is \(Ω(\min\{T-L_0,\sqrt{K(T-L_0)}\})\). It is monotonically decreasing in \(L_0\), even when \(L_0=\min_d L_d\) itself is known. On the positive side, this result motivates the possibility that other coordinates, not just the easy one, may suffice to attain the optimal rate of Pareto regret. When $L_0$ is known, we apply Poly-INF to a fixed coordinate and obtain an upper bound on Pareto regret that exhibits the same order and thus matches the lower bound. Without such knowledge, we develop a reward-doubling version of Poly-INF that adapts to this unknown quantity while still attaining the matching minimax rate. Another implication is that it has no extra \(\log T\) factor and is independent of the number of coordinates.