Mitchell Troyanovsky, co-founder of Basis, told AI Engineer how his firm is trying to make agents do production scale accounting work reliably enough for firms to use end to end.
Why does Codex or Claude Code make you feel 100 times faster while your company shows almost no lift at all?
Troyanovsky argued in the talk that the answer is not smarter models but two gaps enterprises have not closed, a competency gap that pushes complexity onto the user and a process gap that lets every employee decide how the agent should work.
Humans already work with non-deterministic colleagues, he said, and firms scale trust not through deterministic control but through processes handed down from leadership that don't require everyone to be perfect to get consistent outcomes.
A behavior spec is the written intent that captures the most important opinion about how the agent must act, and Basis uses it as a rubric to judge whether a trajectory followed the opinion even when no ground truth answer exists.
The daily brief
Get the free daily brief with the day's AI news, new AI startups and funding rounds every morning.
Join 64,000 other subscribers.
An opinion is a deliberate choice among infinite implementations, such as requiring the agent to read the primary tax source before answering or to photograph a PowerPoint for formatting, which trades a little latency for reliability and must be applied consistently rather than as occasional genius.
To close the loop without waiting for thousands of outcome based evals, Troyanovsky detailed forking, which clones a failed trajectory right before the wrong decision so you can ask why or edit context, and simulations, which spin up a cheap sub-agent that pretends to be production to test tweaks quickly.
Basis, founded in 2023 by Matt Harpe and Troyanovsky to build AI agents for accounting work according to an Apple Podcasts listing, illustrates the shift that finance leads and firm partners must now author opinions centrally instead of leaving customization to each user.
