Atlas Capital · 2025
Autonomous research agents
A fleet of investment‑research agents that read filings, run financial models and draft memos — supervised, not autonomous.
Tag
AI Agents
Duration
6 weeks
Role
Agent design · Eval · Frontend
Memos / analyst / week
3.2× more
Time on bookkeeping
−68%
Hallucination rate
<0.4%
◇ The problem
Where they were stuck.
Junior analysts were spending 70% of their week on bookkeeping work — reading 10‑Ks, building DCFs, formatting memos.
◇ Our approach
How we shipped it.
- 01
Decomposed analyst work into 9 deterministic + LLM-graded sub-tasks.
- 02
Built a tool-using agent with strict schemas, caching and replayable traces.
- 03
Designed a memo authoring surface that lets analysts edit any agent step.
- 04
Hardened with an LLM-judge eval suite over 1,200 historical memos.
◇ Stack
What it's made of.
LangGraphPythonPostgresReactOpenAIAnthropic
Next case study