TRADING 101 / RULES, MODELS & AGENTS
Algorithmic vs agentic trading: what changes?
An algorithm follows a procedure. An AI model supplies an interpretation or estimate. An agent can choose the next step and tool. One trading system can contain all three.
Start with who makes each decision.
In a fixed-rule strategy, the designer chooses the conditions and the program checks them. A model-assisted strategy uses a learned output as one input to a decision. An agentic system can also choose which permitted tools to use as it investigates or acts.
These are overlapping roles. An agent may supervise a conventional rule engine without choosing trades itself. A scheduled script can run autonomously while following the same predefined workflow every time.
Follow one fictional situation.
A coin has risen over the day, but its latest quote is old. These examples show possible designs, not current trading signals.
The rule checks a condition.
“Do not enter using an old quote.” The condition fails, so the program waits. A freshness rule can prevent this particular bad input; it cannot prevent every loss.
The model interprets information.
A learned model might estimate whether the available price history resembles a trend. Its output still needs a defined decision rule and fresh inputs. A confident estimate does not repair stale data.
The agent chooses what to check next.
Given permission, an agent might inspect the timestamp, consult another approved source, and report whether the quote can be trusted. It must wait if required evidence remains unavailable.
What does adding an agent cost?
Tool calls take time and can fail. Models can misinterpret evidence, follow untrusted text, or give inconsistent answers. More decisions mean more behavior to test. Fixed limits on budget, allowed tools and authority should not depend on an agent remembering a request.
Our use of “agent” follows the distinction between predetermined workflows and model-directed tool use described in Anthropic’s explanation of agents. Definitions vary; the label alone tells you little about a trading system.
How would we find out whether it helps?
Compare a simple rule with a separately registered model- or agent-assisted version. Hold the eligible markets, timing, budget and cost assumptions as comparable as possible. Record inputs, model versions, tool failures, delays and decisions. Evaluate later observations that were not used to choose the approach.
Measure useful research assistance separately from trading return. Catching a data error could save work without creating a profitable strategy.
Our deployed paper portfolios follow fixed rules. AI helps develop tools, investigate results and explain evidence. Model-managed and agentic portfolios are future experiments; the examples above do not claim they are running.
Keep exploring.
Try a buy-or-wait decision, inspect the actual BTC, ETH and SOL rules, or learn what paper trading and backtesting can each establish.
For broader context: SEC report on algorithmic trading. Sources explain concepts; they do not endorse CouchChange or validate its strategies.