Why AI Makes You 100x Faster but Not Your Company - Mitchell Troyanovsky, Basis
핵심 요약
Production agents require opinionated design and process supervision to bridge the enterprise productivity gap instead of relying solely on individual coding tools.
하이라이트
Individual knowledge worker productivity increases up to 100x with tools like coding agents, whereas average enterprise productivity remains unchanged.
Scaling organizational productivity relies on established processes and trust rather than deterministic human execution.
Coding-style agent products fail inside companies due to a competency gap that outsources complexity to the user and a process gap that lacks top-down standardization.
Opinionated agent design establishes consistent, reliable execution by deciding specific methods instead of leaving infinite execution paths open.
Behavior specs define intent and act as rubrics to supervise agent trajectories beyond outcome-based evaluations.
Forking agent trajectories right after decision points enables targeted context updates and harness improvements.
타임라인
The Enterprise Productivity Paradox
Founders of Basis focus on making agents perform production-scale work reliably for accounting firms. While individual users experience massive productivity gains with coding tools, companies do not see matching scale. Traditional organizations solve this by implementing standardized processes that head executives trust employees to follow across day-to-day operations.
- Individual developer and knowledge worker productivity multiplies up to 100 times using modern coding agents.
- Average enterprise and firm-level productivity fails to scale by similar orders of magnitude.
- Organizations scale output by trusting employees to follow generalized processes rather than relying on deterministic execution.
The Competency and Process Gaps in Coding Agents
Standard agent tools outsource complexity directly to the user, creating a competency barrier that most knowledge workers cannot navigate. Allowing every employee to invent their own agent workflows violates how companies successfully scale processes. Overcoming these barriers requires making agents simple to use and consistently reliable without demanding perfection.
- Coding-style agent products suffer from a competency gap that forces users to manage agent skills, memories, and theory of mind.
- A process gap occurs when individual employees customize agent behaviors independently instead of following senior leadership standards.
- Successful production agents require simplicity that feels like working with a coworker alongside reliable and consistent execution.
Opinionated Design and Process Supervision
Allowing agents to choose from infinite possibilities for tasks like reading PDFs creates unreliable outcomes. Developers establish opinions based on expertise, such as having tax agents review primary sources to maximize accuracy. Managing agent behavior requires treating runtime context as training data and using behavior specs to evaluate whether necessary actions occur across trajectories.
- Opinionated agent design eliminates infinite execution variations by enforcing specific workflows like reading primary sources before returning answers.
- Runtime training data encompasses every token within system prompts, progressive disclosures, and tool descriptions.
- Behavior specs function as product specs for the agent era to align team intent and supervise trajectories without ground truths.
Forking and Simulations for Agent Improvement
Forking allows developers to investigate why an agent failed a behavior evaluation by inspecting specific decision points within long trajectories. Simulations offer a cheaper alternative by instructing coding agents to roleplay as production agents for rapid context testing. Ultimately, reliable and opinionated agents enable teams to build self-improving systems that scale enterprise output safely.
- Forking copies agent environments right after decision points to diagnose and fix behavior eval failures.
- Simulations utilize coding agents to build sub-agents that mimic production agents for cheap and fast context iteration.
- Diffusion of agents across companies depends on simplicity, reliability, and opinionated behavior that closes the self-improvement loop.
설명
Individual engineers have seen huge productivity gains from AI, but most companies haven't. Mitchell Troyanovsky, co-founder of Basis, argues that agents need to be simple and reliable to spread through a company, and the way to get there is to make them opinionated. He covers behavior specs, why outcome evals don't generalize to production, and two tools for finding out why an agent made a bad call: forking and sub-agent simulations. 0:00 Intro to Basis 0:42 Why company productivity hasn't jumped 1:41 How organizations scale with processes 3:29 The competency gap and the process gap 5:57 Agents need to be opinionated 9:02 Why outcome evals don't generalize 10:18 Example: a bank reconciliation behavior 12:03 Behavior specs 13:24 Forking to diagnose bad decisions 15:15 Simulations with sub-agents 17:10 Earning trust like a coworker 18:10 Why not let agents find a better way? Mitchell Troyanovsky, co-founder of Basis Join us in SF for the second annual AIE CODE summit: https://aie.codes our last flagship summit of 2026. Applications live now!