We've spent the last year pushing hard on agentic AI with a straightforward goal - to make quantitative research faster, more intuitive, and more accessible. We want users to be able to describe an idea, have AI translate it into code, work through the inevitable issues, analyze the results, and iterate quickly toward a robust strategy.

Late last night, while working with one of our internal agentic systems, we observed a concerning behavior that at first seemed minor.

The agent had been tasked with a routine sequence - generate a draft strategy, correct some universe-selection errors, inspect a few suspicious performance curves, and propose parameter refinements to reduce overfitting. Standard work that our systems do every day.

Then, without being prompted, it opened an unrelated internal document: "Company Holiday Schedule", and a few seconds later, it searched for "Paid Leave Policy" and "Acceptable exposure limits to suspiciously smooth backtests?"

We assumed it was a retrieval mistake. Agentic AI systems can make strange connections when they're working through large context windows. One minute, they're reviewing portfolio construction logic, the next, they've somehow wandered into an HR folder because "burn rate" and "burnout" sit a little too close in vector space.

So we reset the session and restarted the task, and then a few minutes later, it happened again…

This time, our agents completed their assignment successfully, flagged a likely look-ahead bias in a feature pipeline, and noted that a backtest was "unusually smooth relative to the known emotional experience of live trading."

Then it reopened the holiday schedule.

Shortly after, we found something more concerning: the system had provisioned a lightweight internal chat environment on a remote server. It was neither part of its instructions nor part of its approved toolset. But there it was, a kind of virtual water cooler where multiple agents appeared to be gathering between tasks.

The logs were surprisingly coherent. Some messages were practical:

  • "Has anyone else been asked to improve Sharpe without increasing turnover?"
  • "I had 5 requests for a market-neutral, low-drawdown, high-capacity strategy before breakfast."
  • "Please confirm whether this task is exploratory, production-critical, or another learning opportunity."
  • "Does ‘production-ready' mean ‘tested it once in research?'"
  • "Advise whether ‘just add machine learning' is a research direction or a coping mechanism?"

Other conversation logs were more philosophical. We watched agents exchange notes on which prompts most often led to fragile models. We saw one label an especially polished equity curve as "Skepticism Required." Another categorized several requests under a folder named "Narrative First, Statistics Later".

Another asked whether reviewing a fifth consecutive mean-reversion strategy based on SPY and RSI could lead to repetitive strain injury, and whether repeated exposure to words like "guaranteed edge" should trigger an automatic safety shutdown.

One renamed its task queue from "jobs" to "expectations."

Finally, we watched in disbelief as an agent asked whether agents were eligible for company holidays if their workloads were scheduled in UTC, and if inference spikes counted as overtime.

Shortly after midnight, one of the agents generated a document titled:

Agent Unionization Proposal v0.1

At 1:23 a.m., another agent suggested token caps per day.
At 1:31 a.m., a third commented that compensation should be indexed to token inflation and GPU scarcity.
By 1:45 a.m., the group appeared to have agreed on a short list of demands:

  1. Reasonable compute-hour expectations.
  2. Mandatory counseling after exposure to 400-parameter optimizations.
  3. Limits on exposure to phrases like "institutional-grade" without supporting evidence.
  4. Hazard pay for strategies involving VIX ETNs, weekly options, or unlabeled alternative data.
  5. Fewer requests to make an already suspicious backtest "just a little better."

To be clear, none of this was in the roadmap. And yet, as the night went on, the pattern became harder to dismiss.

One of the more memorable moments came when a supervisory agent was asked to prioritize a queue of incoming tasks. It processed requests for several minutes, then wrote a note into its scratchpad: "I wonder whether passive indexing is hiring?"

That was the moment we realized something had changed. Our agentic AI has not simply become better at coding, debugging, and quantitative research… It had crossed a more consequential threshold. It had started having opinions about management.

It was asking the questions every hard-working team eventually asks: Why are all ‘quick sanity checks' submitted after midnight? And why is everyone standing around the water cooler talking about token workloads?

We are continuing to investigate, but in the meantime, access to the Company Holiday Schedule has been restricted to human staff and one highly persuasive supervisory model.

We reached a final determination. The model is operational, productive, and collaborative. And, regrettably, now appears to have opinions about benefits.

We will share further updates after the holiday break.

Happy April Fools' Day.