▶ Watch ↗19:12
Conference affiliation: Basis
Mitchell Troyanovsky, known as Mitch, is a co-founder of Basis, which builds AI agents for accounting assignments that can take hours or days, including partnership tax returns. His work addresses a problem that becomes harder as agents take on longer assignments: a plausible result does not reveal whether the agent followed a sound accounting procedure, and engineers cannot supply an answer key for every situation it encounters.
Before Basis, Mitch left Goldman Sachs’s Marquee business to join Puzzle, where he worked on financial accounting software.
Mitch and fellow founder Matthew Harpe introduced the Basis platform in October 2023, alongside a $3.6 million seed round. They developed it with accounting firms facing growing workloads and staffing constraints, aiming to place agents inside existing workflows. Basis’s products span tax preparation, client accounting, audit and corporate accounting, with accountants reviewing the work agents produce.
That shift requires changes within accounting firms as well as better software. Mitch co-authored the April 2025 introduction of Basis’s Deployed Intelligence team with Nikhil Lohchab and Luke Bradley. The team works inside customers’ organizations to understand their circumstances and redesign workflows around agents. Accountants accustomed to performing each step must learn how to direct and review delegated work, while the implementation must accommodate established practices and older software.
Mitch identifies two obstacles to turning individual productivity gains into organizational improvements: the competency gap and the process gap. Each user of a general-purpose agent must learn to manage its context and capabilities, then decide how it should perform the firm’s work. An accounting firm already has procedures for tasks such as extracting bank data; letting every employee independently configure an agent can undo that shared approach.
Mitch examines a related problem inside Basis in his August 2025 essay on building a company for the AGI era. Access to a capable model still leaves employees assembling the information it needs. Even adding a dashboard filter requires knowledge of the repository, interface conventions and database terminology. Reconstructing that context for each request duplicates work across the company.
His proposed response is to treat company context like a codebase: assign owners, version changes, review updates and manage deployment. Basis established its Atlas team to maintain that knowledge, build integrations and develop internal agents. Mitch described this as work in progress, requiring the organization to adapt alongside its software.
In July 2026, Mitch and Braintrust co-founder Ankur Goyal published behavior specs, an open standard developed through the two companies’ collaboration. The approach separates a team’s expectations for an agent from the prompts, tool descriptions and other context used to produce its behavior.
A spec records an important recurring decision without duplicating every instruction. Reviewers can then ask whether a situation called for that behavior and whether the agent exhibited it. This adds evidence about runs for which no complete answer key exists; it supplements outcome evaluations without proving that every result is correct.
Mitch’s accounting examples make the distinction concrete:
These expectations involve tradeoffs. Having an agent inspect an image of a presentation can catch formatting problems but adds latency. A team may decide that the extra check is unnecessary for a sufficiently low error rate. The purpose of a behavior spec is to make the chosen expectation explicit and testable.
Mitch also describes two ways to investigate failures. Forking recreates a run around a questionable decision so engineers can test changes to the agent’s context, although reproducing its database and environment can be expensive. Simulations use a sub-agent supplied with the production agent’s instructions and skills to explore the same scenario. He considers simulations less faithful but cheaper and faster, and says he used them for more than a year before Basis built internal forking tools. Together, these methods help engineers investigate the decisions behind a failure rather than infer its cause from the final result alone.
▶ Watch ↗19:12
Loading more talks and highlights…