Market
Tech Due Diligence in the age of agents: what Brazilian investors should check in 2026
Brazilian PE and VC funds still treat Tech DD as a compliance checklist. When the target runs on AI agents, that checklist misses 70% of the relevant risks.
The first generation of Tech DD in Brazil copied the Big Four playbook: check whether source code is versioned, tests are automated, infra is on cloud, there's a DR plan. That's still necessary. But in AI-first targets, it's far from sufficient.
What changes when the operation depends on an agent
A vertical AI agent has operational risk different from a traditional stack. Output depends on third-party models (Anthropic, OpenAI, Google) — changes in those models change the product's behavior. If a firm depends on LawPilot to triage 50k cases per month and MVA swaps the base LLM, the firm feels it.
Assessing that risk requires looking at evals: what does the company measure to guarantee agent quality? How often does it run? Is there automated rollback?
Five specific checks for AI-native targets
1. LLM dependency model: who is the primary vendor? Is there a fallback? What's the cost of switching? 2. Evaluations: is there an automated eval suite running on PRs? Coverage? Historical drift tracking?
3. Cost control: what does a single execution cost? What's the margin above inference cost? Compressing trend? 4. Context engineering: how is the agent kept with regulatory context? Humans curating prompts? A proprietary agent? 5. Observability: can you audit a specific execution? Are reasoning logs stored?
How MVA structures that scoring
Our framework has 26 dimensions grouped into 4 categories (architecture, team, operations, security). For AI-first targets, we extend the 'operations' module with 5 additional dimensions: model risk, eval discipline, inference economics, context pipeline, behavioral observability.
Concrete result: in 2025, in three PE deals reviewed by MVA, those five dimensions flagged material risks that a traditional checklist would have missed. In two of those deals, term sheet conditions changed.
Other posts
View allThesis · 6 min
The post-SaaS era: why the next generation of legal tech won't look like SaaS
Brazilian legal tech was built on SaaS logic: sell better software so lawyers do the same work with less friction. What changes when AI starts to do the lawyer's work?
Read postPlaybook · 8 min
Anatomy of a multi-agent architecture for judicial auction monitoring with Claude + AWS
How Leiloa.ai processes 12 thousand judicial auction notices per month with an ensemble of agents that uses Claude for semantic extraction and AWS Lambda for orchestration.
Read post
