Engineering Autonomous Agents for Due Diligence

The marketing is full of "autonomous AI agents" that promise to transform due diligence. Most are chatbots with API access.

Real autonomous agents — ones that break down complex analysis tasks, run multi-step workflows, and self-correct when they hit ambiguity — take serious engineering. This is systems design, not prompt engineering. Here's what separates real agents from demos.

Andrej Karpathy said in December 2025: "I don't think the industry has realized anywhere near 10% of their potential even at present capability." Due diligence is one of the clearest places that potential is sitting unused.

"The biggest advantage of LLMs for coding is not getting work done faster but being able to ship projects that wouldn't have been justified spending time on at all."

Simon Willison, software engineer and AI researcher (March 2025)

The Architecture Challenge: Beyond Single-Model Inference

Building an autonomous agent for due diligence isn't about picking the right foundation model. It's about orchestration. One LLM call can pull data from a CIM. But reading a 200-page memo, comparing it to industry benchmarks, flagging EBITDA adjustments, and writing an investment committee memo takes several specialized agents working together.

That creates three engineering problems: state management (how do agents share context without memory leaks?), tool orchestration (how do you chain function calls without cascading failures?), and error handling (what happens when an agent hits a broken Excel file or vague legal language?).

Three Critical Design Patterns

1. Break the work into specialist agents. Instead of one monolithic "due diligence agent," our Pre-Screening Agent uses several: a financial agent for EBITDA normalization, a legal agent for regulatory risks, an ESG agent for impact screening. Each one has a narrow job and clear success criteria. The orchestrator routes subtasks to the right specialist.

2. Escalate to humans when confidence drops. Autonomous doesn't mean unsupervised. Every agent output includes a certainty rating. When certainty drops below a threshold — say 85% on an EBITDA adjustment — the system kicks it to a human instead of guessing. This avoids the "95% accuracy problem": in high-stakes finance, 5% errors can cost millions.

3. Self-correction loops. Real agents need to fix themselves. If the financial agent hits missing data in a revenue table, it doesn't fail silently. It flags the gap, logs it, and keeps processing other sections. Analysts fill in the missing inputs without restarting the whole workflow.

Real-World Constraints: Security, Cost, Latency

Data Security. Due diligence involves highly confidential data. Our data is never stored after processing — never used to train public models.

Cost Management. Multi-agent systems multiply costs fast. A naive build might make 50+ LLM calls per deal. We use caching, selective tool calls, and smaller models for simple subtasks to keep costs predictable.

Latency. A screening memo that takes 8 hours isn't autonomous — it's a batch job. Real-time orchestration with parallel execution keeps our average processing time under 4 hours.

Building autonomous agents for due diligence is closer to distributed systems design than prompt engineering. Task breakdown, verification checkpoints, and self-correction aren't optional features. They're requirements. Firms that get this will deploy agents that scale. The rest will deploy demos.