Basis

"It's going to sound silly, but I promise it's very effective."

@mitch_troy's favorite debugging trick: have a subagent pretend to be your prod agent. Tell it not to cheat. Replay the scenario and tweak the context.

"You can do this today without building anything."

@aiDotEngineer NYC:

AI made engineers much faster. Most companies didn't follow.

@mitch_troy at @aiDotEngineer NYC on why: agents spread when they're simple and reliable, and reliability comes from opinions. Plus behavior specs over outcome evals, and debugging bad calls with forking and sub-agent simulations.

19 min:
getbasis.ai/blogs/why-ai-m…

"It failed the outcome. How do you know why? You have potentially a thousand inference calls in your 5 hour long agent trajectory."

@mitch_troy on forking: rewind to right before the bad decision, fork it, and ask the agent why it did that.

From @aiDotEngineer NYC:

Introducing Basis in your binder.

Basis now runs inside the audit binder your firm already uses. Agents work right alongside your team, and auditors start at review.

No moving documents between tools. No writing prompts. Your firm's standards apply on every engagement.

Firms that start now head into busy season with agents already doing the first pass.

Read the full post → getbasis.ai/blogs/introduc…

"What you don't want is a customer getting one way one time and another way the other time."

@mitch_troy on behavior specs: "a product spec for the agent era." Write down the opinions that matter most, then use them as a rubric, even with no ground truth.

@aiDotEngineer NYC: