Agents Need Workflows

If you’ve tried ‘AI agents’ in a real company, you’ve probably seen the same pattern: the demo looks magical, then the agent meets messy inputs, ambiguous goals, and unclear definitions of done—and the whole thing turns into a polite chaos machine.

The practical takeaway for founders is uncomfortable but empowering: agent performance is usually capped by workflow design, not model IQ.

A recent arXiv paper on multi-agent LLM systems for trading makes this unusually concrete. The authors argue that many multi-agent systems imitate an investment team with broad roles (analyst, manager), but those designs often rely on coarse instructions that don’t reflect real-world workflow complexity.

They propose explicitly decomposing the job into fine-grained tasks and evaluate it in a leakage-controlled backtest using Japanese stock data and multiple input sources (prices, financial statements, news, macro signals). They report improved risk-adjusted returns compared to coarse-grained designs.

The thesis for founders: stop building ‘role agents,’ start building ‘task lines.’ Each step should take bounded inputs, produce an auditable artifact, and have a definition of done—plus explicit stop conditions.

Trading is just a harsh environment where weak systems get punished quickly. But the failure modes map cleanly to everyday company workflows: underspecified objectives, messy inputs, implicit quality bars, and no intermediate artifacts you can inspect when it goes wrong.

What to do next week: pick one high-frequency workflow and decompose it into 6–10 micro-tasks (not roles). For example, inbound lead qualification can be split into extraction, summarization, disqualifiers, rubric scoring, draft response, next-step recommendation, and logging.

Add checkpoints and stop conditions: human approval before external sends or state changes, and ‘stop and ask’ behavior when required fields are missing or confidence is low.

Instrument it like a product: completion rate, human edit distance, time-to-done, and failure modes by step. Step-level visibility is what lets you iterate toward reliability.

The real moat isn’t prompts. It’s legibility plus feedback loops. You don’t need smarter agents—you need clearer work.