Casper Dong

on the atlas — 3

• Autoresearch Is Not About Training Models. It Is About What Happens When Agents Get a Scoreboard | by Krish | Mar, 2026 | AI Sutra1 savers
• [2603.16862] Chronos: Temporal-Aware Conversational Agents with Structured Event Retrieval for Long-Term Memory1 savers
• Curius / Onboarding2621 savers

highlights — 9

• The constraint is having a clear metric and a well-structured codebase.

  Autoresearch Is Not About Training Models. It Is About What Happens When Agents Get a Scoreboard | by Krish | Mar, 2026 | AI Sutra

• There is also the question of what happens when agents start modifying codebases that interact with external systems. Optimizing a self-contained training script is one thing. Optimizing a microservice that talks to a database, a cache, and three other services is a different problem entirely. The blast radius of a bad experiment goes up.

  Autoresearch Is Not About Training Models. It Is About What Happens When Agents Get a Scoreboard | by Krish | Mar, 2026 | AI Sutra

• Reviewability is another concern. 93 automated commits from 120 experiments means someone still needs to review what the agent did. The diffs are human-readable because autoresearch constrains modifications to a single file. But as the pattern scales to larger codebases, maintaining that reviewability becomes harder.

  Autoresearch Is Not About Training Models. It Is About What Happens When Agents Get a Scoreboard | by Krish | Mar, 2026 | AI Sutra

• I would not be doing my job if I did not flag the risks. Overfitting to the benchmark is the obvious one. Tobi Lutke himself flagged this when he shared the Shopify results. If your evaluation metric does not capture what actually matters in production, the agent will optimize for the wrong thing enthusiastically. This is the Goodhart’s Law problem applied to autonomous agents, and it is real.

  Autoresearch Is Not About Training Models. It Is About What Happens When Agents Get a Scoreboard | by Krish | Mar, 2026 | AI Sutra

• The multi-agent version is an engineering problem at this point, not a research problem. Someone will build it. Probably several someones. And when they do, the gap between organizations that participate in that network and those that do not will widen quickly.

  Autoresearch Is Not About Training Models. It Is About What Happens When Agents Get a Scoreboard | by Krish | Mar, 2026 | AI Sutra

• Distributed agents, each running experiments, sharing findings, building on each other’s work.

  Autoresearch Is Not About Training Models. It Is About What Happens When Agents Get a Scoreboard | by Krish | Mar, 2026 | AI Sutra

• Consider what Tobi Lutke did. Shopify’s CEO pointed an autoresearch-style loop at Liquid. For those who do not know,

  Autoresearch Is Not About Training Models. It Is About What Happens When Agents Get a Scoreboard | by Krish | Mar, 2026 | AI Sutra

• Chronos applies dynamic prompting to generate tailored retrieval guidance for each question, directing the agent on what to retrieve, how to filter across time ranges, and how to approach multi-hop reasoning through an iterative tool-calling loop over both calendars.

  [2603.16862] Chronos: Temporal-Aware Conversational Agents with Structured Event Retrieval for Long-Term Memory

• Starting [at a young age] he’s read everything that he

  Curius / Onboarding