Mitchell Troyanovsky
Co-founder at Basis
About Mitchell Troyanovsky
Mitchell Troyanovsky is one of the two co-founders of Basis, a New York startup building AI agents that do accounting, tax and audit work for accounting firms. He founded the company in 2023 with CEO Matt Harpe. Before Basis, he was the first product hire at Puzzle Financial, an accounting ledger startup. Basis reached a $1.15 billion valuation in February 2026.
In an AI Engineer talk, Troyanovsky explains why agents that make individuals 100x more productive haven't done the same for companies, and how Basis builds agents firms trust end to end.
Career Highlights
- Co-founder of Basis (2023), AI agents for accounting firms
- First product hire at Puzzle Financial
- Speaker at AI Engineer (2026) on opinionated, reliable production agents
Notable Positions
On the Competency and Process Gaps
Coding-style agents push complexity onto users, who must manage skills and memory, and let each user decide how work gets done. Firms scale through processes set from the top, so Troyanovsky argues production agents should feel like a co-worker and follow the firm's process.
On Opinions and Behavior Specs
Reliability comes from opinions about how the agent should work, such as reading the primary source before answering a tax question. Basis writes behavior specs that list the opinions that matter most and uses them as rubrics for trajectories that have no ground truth. See Agent Evaluation.
On Forking and Simulation
To learn why an agent broke a behavior, Basis forks the trajectory at the decision point. The cheaper version is asking a coding agent to simulate the production agent with the same prompt and skills.
Key Quotes
- "Humans are already used to working with non-deterministic systems. It's just those systems are normally their co-workers, not their computers." (on how firms scale)
- "Nobody wants their agent to move 37 their tax return." (on why reliability beats novelty)
Related Reading
- Agent Evaluation - Testing agents for production work
- Context Engineering - Treating context as runtime training data