What an AI trading agent does
Give the agent an objective: follow a company, compare a watchlist or manage a trading assignment. It uses tools to investigate, reads the results and decides which research step to take next. It returns evidence and reasoning you can review.
Premiss agents use web research, SEC filings, historical market data and saved findings. Schedule recurring tasks and receive reports in chat, your Desk or connected Discord destinations. Trading permissions determine whether an agent proposes an action or carries it out.
Two ways to reach a trading decision
A fixed strategy applies explicit conditions, such as checking whether Bitcoin has crossed a moving average at a completed daily close. An AI agent can instead weigh filings, market data and your brief to decide whether to take a position. Its conclusion may change as the evidence changes.
Premiss supports both approaches. For a fixed strategy, inspect the rules and simulated trades. For an agent’s decision, inspect the research, reasoning and proposed action. An explanation helps you review a decision; it does not make that decision reproducible as a trading rule.
Choose how much work the agent handles
Choose the outcome for each task. A weekly stock briefing and an automatic crypto trading assignment need different instructions, even if they use some of the same research sources.
- Research: gather evidence, compare sources and report the findings.
- Review: explain the proposed action and bring it back for your decision.
- Automatic trading: take positions within the markets and limits you set.
Make trading permissions concrete
Choose the trading account or destination, permitted instruments, position sizing and exposure limits. State when the agent can act and when it must bring a proposal to you. Avoid instructions such as “take sensible risks,” which leave those decisions undefined.
Configure the controls in the service that handles the orders as well. Agent instructions describe intended behavior; account settings determine which orders a connection accepts. Check execution records: sending a signal, receiving it and filling an order are separate events.
Decide how to monitor open positions and pause the agent. Check what pausing does to orders already sent and positions still open in the execution account.
Evaluate fixed rules and changing decisions differently
For a fixed strategy, run a historical backtest. Inspect the generated code, data period, costs and individual trades. Check whether the same rules behave differently across market conditions or when you change an assumption, then observe the strategy in paper trading as new data arrives.
An agent that weighs new evidence needs a record of decisions made as that evidence becomes available. Start with paper decisions. For each one, keep the source dates, reasoning, task instructions and action. Review that record over time. Running today’s agent over old prices while allowing today’s filings or web results cannot show what it could have known then.
Check missing data, decisions to do nothing and differences between intended and simulated positions. Keep changes to the agent’s brief on record so you know which instructions produced each decision. Backtests and paper results remain hypothetical; they do not guarantee future returns.
Use feedback to keep the agent aligned
Review reports for unsupported conclusions, missing context and repeated assumptions. Ask for sources and dates. When you correct a finding or change your objective, update the task so the next run has an explicit brief.
Compare trading decisions with the brief, supporting evidence and execution records. Reconsider the agent’s limits when you change its assignment, and use ongoing review to keep the work relevant.