Tauric Research
Perspectives

Multi-agent orchestration

  • Foundation

What multi-agent systems add, what they cost, and when they pay off.

Investment research is usually done by teams: an analyst builds the case, a colleague argues against it, and a risk manager asks what happens if both are wrong. Multi-agent language model systems try to reproduce that structure. The evidence on when it helps is mixed, and the mix is instructive.

State of the field

Anthropic reports that its multi-agent research system outperformed a single agent by 90.2% on an internal research evaluation, while using about 15 times more tokens than a chat. It also notes that domains where all agents must share the same context, or where agents depend heavily on one another, “are not a good fit for multi-agent systems today” (Anthropic, 2025 (external site)).

A controlled study from Google Research makes the trade-off concrete. On a financial reasoning task that splits into parallel parts, centralized coordination improved performance by 80.9% over a single agent; on tasks that require strict sequential reasoning, every multi-agent variant tested degraded performance, by 39 to 70%. Agents working in parallel without communicating amplified errors 17.2 times, against 4.4 times when an orchestrator coordinated them (Kim and Liu, 2026 (external site)).

Investment work has a natural division of labor to borrow from. TradingAgents (external site) (Xiao et al., 2024) models a trading firm, with fundamental, sentiment, news, and technical analysts, bull and bear researchers, a trader, and a risk team who debate a decision before it is made.

Principles

Specialization pays when an analysis breaks into distinct perspectives, as many investment decisions do. When the reasoning is one long chain, extra agents add cost and new ways to fail. Anthropic’s guidance is to find “the simplest solution possible” and add complexity only when it is needed (Anthropic, 2024 (external site)). Whatever the structure, something should check the agents’ work, and their disagreements should be kept, because they are often the most useful part of the research.

Our approach

TradingAgents, our published framework, divides the analysis of a trade among agents in defined roles. Our design direction builds on it: each agent has a specialization and manages its own context and memory, agents coordinate through defined protocols, work is decomposed into units small enough to run and check, and a workflow graph routes and coordinates the steps.

References